Perception Coverage Audit and Backlog
The perception library is broad, but the May 2026 agent sweeps found several method families that are still hidden inside larger documents or missing as first-class pages. Treat this file as the backlog for turning perception research into a method-level library similar to the SLAM coverage audit.
Current Status
| Item | Status |
|---|---|
| Dedicated perception files | 38 top-level perception synthesis/audit files, 36 dataset/benchmark pages, and 137 atomic method files plus the method-library overview in Perception Method Library. |
| Strongest existing coverage | Production sensor suites, CenterPoint/OpenPCDet, BEV basics, method-level camera BEV/occupancy including DepthOcc, sparse-query camera detection, sparse cooperative queries, collaborative Gaussian semantic occupancy, LiDAR-camera occupancy fusion, dynamic occupancy/free-space, spatiotemporal occupancy flow, Gaussian/3DGS/4DGS perception including GS-Occ3D label curation, LiDAR MOS, scene flow, LiDAR artifact removal, learned denoising, adverse-weather datasets, truck-centered V2X datasets, embodied 3D perception benchmarks, FOD synthetic benchmarks and Airport-FOD3S data-engine workflow, motion/static separation benchmarks, aggregated-map semantic segmentation with compact proxy/input/training selection plus companion pages for class taxonomy, training paradigms, tiling/throughput, post-processing, GridNet-HD LiDAR-image utility-infrastructure proxy routing, point-cloud Mamba/SSM backbone routing, LOSC open-vocabulary LiDAR label consolidation, map-hygiene ground-truth workflow routing, static-but-transient point removal, detector-based potentially dynamic object removal, and SLAM map-version-control handoff, radar-camera/radar-LiDAR fusion, 4D radar-camera occupancy, 4D radar road-boundary/freespace detection, radar/event/FMCW, open-world/OOD, open-vocabulary attributes, panoptic/open-vocabulary occupancy, proposal-free open-vocabulary 3D instance segmentation, robust fusion, V2X, latency, data engines, model compression, uncertainty, and thermal fusion. |
| Most severe structural gap | Remaining P0/P1 items still need atomic pages or routing: cooperative semantic-communication and collaborative Gaussian follow-ons beyond VOGS-CP, source-mature LinkOcc or related temporal occupancy follow-ons, occupancy uncertainty/OOD follow-ons beyond the promoted EvOcc page, and public-data gaps for dust, de-icing mist, steam, glycol film, and wet-apron multipath. SAM4D already has an atomic page and should stay routed as a promptable candidate-label/fusion method rather than a missing severe gap. The FOD/adverse-airside proxy boundary is now explicit across dataset, validation, and removal pages; direct airside collection remains the unresolved gap. DR-REMOVER and ExelMap are covered in localization/mapping pages rather than missing perception pages. |
| How to use this audit | Use the method library for promoted atomic pages. Use the P0/P1/P2 rows below as discovery clusters and backlog for remaining splits. Update this audit whenever a missing perception method becomes a dedicated file. |
Multi-Agent Discovery Sweep (2026-05-08)
Six parallel research agents audited camera/BEV perception, LiDAR/radar/thermal perception, foundation/open-world perception, temporal tracking, cooperative/V2X perception, and deployment robustness. A second targeted sweep then checked 2025-2026 latest-method and benchmark gaps, followed by lightweight dataset scouts for adverse-weather, event, radar, FMCW LiDAR, V2X, OOD, and latency benchmarks.
Method-Level Promotion Wave (2026-05-08)
Five parallel writing agents converted the highest-value discovery clusters into atomic method pages under methods/. Existing broad perception files should now link into this library rather than expanding into larger catch-all documents.
| Cluster | Promoted method files |
|---|---|
| Camera BEV and occupancy | BEVDet, BEVDepth, BEVStereo, SOLOFusion, TPVFormer, SurroundOcc, SparseOcc, FlashOcc, SelfOcc, RenderOcc |
| LiDAR motion, radar, event, and FMCW | LiDAR-MOS, 4DMOS, InsMOS, StreamMOS, 4DSegStreamer, SegNet4D, Mask4D, Instantaneous Motion Perception, RadarPillars, K-Radar, V2X-Radar, Ev-3DOD, AevaScenes |
| Open-world and open-vocabulary | OpenAD, OP3Det, WildDet3D, DetAny3D, OW-OVD, Clipomaly, S2M, SAM 3, 3D-AVS, Mosaic3D, SpaCeFormer, OpenVox, LOSC |
| Robust fusion, calibration, and validation | MoME, GraphBEV, SOAC, RC-AutoCalib, ASF, MSC-Bench, MultiCorrupt, S2R-Bench, Occluded nuScenes, Conformal Boxes |
| Cooperative, latency, closed-loop, and data engines | RCooper, HoloVIC, CoInfra, V2X-ReaLO, CoHFF, CoSDH, CoopTrack, LASP, Fail2Drive, AIDE |
Deep-Dive Promotion Wave (2026-05-09)
Fresh parallel discovery and writing agents promoted the first Gaussian/4DGS and latest radar-camera/sparse-query gaps into atomic pages. The rows below are now the primary pages for those methods; older backlog rows remain as provenance for related follow-ons.
| Cluster | Promoted method files |
|---|---|
| Gaussian and 3DGS perception | SplatAD, GaussianFormer, GaussianOcc, Streaming Gaussian Occupancy |
| 4D occupancy and sparse-query perception | Cam4DOcc, StreamingFlow, Sparse4D |
| Radar-camera fusion | TacoDepth, RaCFormer |
LIORNet and Adverse-Weather Removal Wave (2026-05-09)
The next loop broadened the LIORNet question from one snow-removal method into a full removal stack: learned LiDAR denoising, classical outlier filters, weather artifact removal, ghost/multipath failure modes, dynamic map cleaning, validation evidence, and weather robustness datasets.
| Cluster | Promoted files |
|---|---|
| Learned LiDAR denoising/removal | LIORNet, LiSnowNet, SLiDE, TripleMixer, 3D-KNN Blind-Spot Desnowing, 3D-OutDet, AdverseNet, DenoiseCP-Net |
| Classical and broad artifact removal | Classical LiDAR Outlier Removal, LiDAR Weather Artifact Removal, LiDAR Artifact Removal Techniques, LiDAR Ghost and Multipath Artifacts |
| Weather robustness datasets | Weather Robustness Datasets, WADS, CADC/CADC+, SemanticSTF, REHEARSE-3D, RainSense, SemanticSpray, RADIATE, DSERT-RoLL, CMHT, Seeing Through Fog/DENSE |
| Safety and map-cleaning bridge | LiDAR Artifact Removal Validation, ERASOR, Removert, LiDAR Map Cleaning and Dynamic Removal |
Dynamic/Static Removal and Flow Benchmark Wave (2026-05-09)
The latest loop broadened removal from "delete noisy LiDAR points" into map hygiene: dynamic objects, static objects that do not belong in the persistent map, moved objects, scene flow, moving/static separation, and validation against false deletion.
| Cluster | Promoted files |
|---|---|
| MOS and scene-flow methods | MotionSeg3D, MambaMOS, Neural Scene Flow Priors, Scene Flow for Dynamic Object Removal |
| Motion/static and 4D occupancy benchmarks | Moving/Static Separation MOS Datasets, Occupancy-Flow and 4D Occupancy Benchmarks, Scene-Flow Datasets and Benchmarks |
| Map-cleaning validation bridge | Dynamic Map Cleaning Benchmarks, Moved-Object and Map-Change Datasets, Airside Dynamic Map-Cleaning Benchmark |
Web Gap Expansion Wave (2026-05-09)
Five web-search scouts and six writing agents promoted a new batch of radar, neural-field, dataset, and validation gaps into first-class pages.
| Cluster | Promoted files |
|---|---|
| Radar, 4D radar, and FMCW perception | CVFusion, 4D Radar-Camera Occupancy, POD FMCW LiDAR Predictive Detection |
| Dynamic Gaussian, neural-field, and temporal occupancy | DrivingGaussian, HUGS, SplatFlow, DistillNeRF, TrackOcc, Cross-Domain LiDAR Scene Flow |
| World-model bridge | Self-Supervised Occupancy Flow, UniScene Occupancy-Centric Generation |
| Datasets and validation | MUSES, Sensor Corruption Robustness Benchmarks, Open-World OOD and Anomaly Segmentation Benchmarks, FOD and Airport Apron Detection Datasets, FOD Perception Validation, Knowledge-Base Evaluation Protocol |
Reliability and Open-World Gap Wave (2026-05-09)
Six fresh web-search scouts and six writing agents promoted another set of non-duplicate perception gaps into first-class files.
| Cluster | Promoted files |
|---|---|
| Occupancy, freespace, and adverse-weather fusion | LiDAR-Camera Occupancy Fusion, Dynamic Occupancy and Freespace, Spatiotemporal Memory Occupancy Flow, 4D Radar Road Boundaries and Freespace, Adverse-Weather Radar-LiDAR 3D Detection, RobuRCDet, SAMFusion, GaussianFlowOcc, GaussTR |
| Open-world and open-vocabulary coverage | OVAD/OVODA Open-Vocabulary 3D Attributes, Open-Vocabulary Panoptic Occupancy |
| Dataset and benchmark coverage | STU 3D LiDAR Anomaly Segmentation, Airside FOD Synthetic Multimodal Benchmarks, RCP-Bench Cooperative Corruption Robustness, V2X Large-Range Sequential Datasets |
Recurring Research Loop Promotion (2026-05-22)
Five parallel research scouts checked 2024-2026 perception, localization, dataset, and foundation gaps. The bounded promotion for this loop added one missing camera occupancy method page and kept the rest of the scout findings in the queue.
| Cluster | Promoted files |
|---|---|
| Depth-aware camera occupancy | DepthOcc |
| Sparse-query camera detection | SparseBEV, DETR4D |
| Occupancy follow-ons | TEOcc, GaussRender |
| Adverse-weather multimodal datasets | DSERT-RoLL, CMHT Autonomous Dataset |
Recurring Gaussian Occupancy Promotion (2026-05-23)
Five parallel read-only scouts checked the active perception occupancy queue. This bounded pass promoted two source-mature Gaussian occupancy methods and corrected LinkOcc routing so the RobuRCDet arXiv ID is not used as a LinkOcc source.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Sparse weakly supervised Gaussian occupancy | GaussianFlowOcc |
| Foundation-model-aligned open-vocabulary Gaussian occupancy | GaussTR |
| Temporal occupancy source hygiene | LinkOcc remains a watchlist item with IEEE/DBLP source metadata; do not classify it as radar-camera without primary-source evidence. |
Recurring GS-Occ3D Promotion (2026-05-23)
Five parallel read-only scouts checked the perception occupancy queue again. This bounded pass promoted GS-Occ3D as a vision-only Gaussian-surfel occupancy reconstruction and label-curation page, while keeping LinkOcc on watchlist because open method details or code were still not available.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Vision-only Gaussian occupancy reconstruction | GS-Occ3D |
| Temporal occupancy source hygiene | LinkOcc remains watchlist-only; do not use RobuRCDet's arXiv ID as a LinkOcc source. |
Recurring Cooperative Compression Promotion (2026-05-23)
Five read-only scouts checked cooperative compression and quantized V2X perception against existing CoSDH, CoopTrack, V2X-ReaLO, and robust-cooperation coverage. This bounded pass promoted QuantV2X because primary sources provide a paper, project page, official repository, codebook-message design, PTQ workflow, and deployment-oriented latency evidence.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Quantized cooperative perception | QuantV2X |
Recurring Sparse Cooperative Query Promotion (2026-05-23)
Five read-only scouts checked the remaining cooperative perception queue after QuantV2X. This bounded pass promoted SparseCoop because the paper is accepted in AAAI 2026, has arXiv and proceedings records, publishes an official MIT-licensed repository, and reports detection, tracking, transmission-cost, and FPS metrics on V2X-Seq and Griffin.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Sparse cooperative detection and tracking | SparseCoop |
Recurring Truck-Centered V2X Dataset Promotion (2026-05-23)
Five read-only scouts checked the remaining cooperative perception queue after SparseCoop. This bounded pass promoted TruckV2X because the dataset has an IEEE RA-L 2025 record, arXiv record, official project page, and Hugging Face dataset release, while CoDS still lacks code/project artifacts and JigsawComm has only a placeholder repository with no released implementation.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Truck-centered cooperative perception dataset | TruckV2X |
Recurring Collaborative Gaussian Occupancy Promotion (2026-05-23)
Five read-only scouts checked the remaining cooperative perception queue after TruckV2X. This bounded pass promoted VOGS-CP because it has an AAAI 2026 proceedings record, arXiv record, official project page, and public repository, and because it cleanly separates collaborative Gaussian semantic occupancy from CoHFF, SparseCoop, and QuantV2X.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Collaborative Gaussian semantic occupancy | VOGS-CP |
Recurring FOD Synthetic Data-Engine Promotion (2026-05-23)
Five read-only scouts checked the FOD/adverse-data queue against existing FOD-A, Airport-FOD3S, RDD5000, weather-proxy, and data-engine coverage. This bounded pass promoted Airport-FOD3S as a focused synthetic FOD data-engine page, added DualFOD/FOD-UAS to the FOD benchmark routing, and added LIDAROC as a LiDAR cover-contamination proxy while keeping airside dust/de-icing/steam/glycol/wet-apron public-data gaps explicit.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Synthetic FOD data engine | Airport-FOD3S Synthetic FOD Data Engine |
| FOD multimodal benchmark routing | Airside FOD Synthetic and Multimodal Benchmarks now includes DualFOD/FOD-UAS and data-engine cross-links. |
| Adverse-weather proxy routing | Weather Robustness Datasets now routes LIDAROC as a sensor-cover dust/contamination proxy, not a direct jet-blast or de-icing dataset. |
| FOD proxy-boundary refresh | FOD and Airport Apron Detection Datasets, Airside FOD Synthetic and Multimodal Benchmarks, Weather Robustness Datasets, FOD Perception Validation, FOD Retention Map-Cleaning Safety Case, and LiDAR Artifact Removal Techniques now separate public proxy screening from target-airside acceptance evidence and add do-not-delete hazard retention wording. |
Recurring Embodied 3D Benchmark Promotion (2026-05-23)
Five read-only scouts checked embodied 3D perception benchmarks, sparse/e2e driving follow-ons, GSCOOP source maturity, active source-registry freshness, and radar/Gaussian SLAM candidates. This bounded pass promoted EmbodiedScan/MMScan as an embodied 3D benchmark page because primary sources provide a CVPR 2024 paper, NeurIPS 2024 Datasets and Benchmarks record, official project pages, a shared official repository, and benchmark tasks.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Embodied 3D perception benchmarks | EmbodiedScan and MMScan Embodied 3D Benchmarks |
| Transfer caveat | The new page treats indoor RGB-D and grounded-language benchmarks as spatial-grounding references, not direct outdoor apron safety evidence. |
Recurring Open-Vocabulary 3D Segmentation Promotion (2026-05-23)
Five read-only scouts checked DySS, sparse/e2e perception follow-ons, cooperative perception, and indoor open-vocabulary 3D instance segmentation. This bounded pass promoted SpaCeFormer because it has an NVIDIA project page, arXiv record, and released WarpConvNet source, while DySS is already routed through the sparse-query overview and remains code-watch rather than a duplicate atomic page.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Proposal-free open-vocabulary 3D instance segmentation | SpaCeFormer |
| Sparse-query routing caveat | DySS remains a sparse-query and streaming-temporal overview item until official code or broader adoption justifies a standalone method page. |
Recurring Cooperative Source-Hygiene Refresh (2026-05-23)
Five read-only scouts checked cooperative/infrastructure source drift after CoInfra, HoloVIC, and V2X-ReaLO had already been promoted. This bounded maintenance pass corrected CoInfra from the RCooper arXiv ID to 2507.02245, corrected V2X-ReaLO from an unrelated arXiv ID to 2503.10034, corrected HoloVIC CVF links to the Ma_HoloVIC CVPR 2024 record, and refreshed routing for V2XScenes, UrbanIng-V2X, V2XPnP, and CoopScenes in the infrastructure-cooperative and large-range dataset pages.
Recurring Point-Cloud SSM Backbone Promotion (2026-05-23)
Five read-only scouts checked the aggregated-map segmentation queue after the Uni-Mapper pass. This bounded pass promoted point-cloud Mamba/SSM backbones as one method-family page rather than separate immature pages per variant, because the family is relevant to map-scale token cost but still lacks production map-release evidence.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Point-cloud Mamba / SSM segmentation backbones | Point-Cloud Mamba / SSM Backbones |
| Aggregated-map routing | Aggregated-Map Semantic Segmentation, LiDAR Semantic Segmentation, and Large-Scale 3D Segmentation Tiling and Throughput now treat SSMs as a research-stage efficiency branch, not a P1 production default. |
Recurring Aggregated-Map Architecture Routing (2026-05-24)
This bounded pass kept existing method pages as the atomic deep dives and added an architecture reading path in the method library, because the aggregated-map hub already compares the families but readers still needed a product-oriented route across the individual method pages.
| Cluster | Refreshed coverage |
|---|---|
| Method-library routing | Perception Method Library Overview now groups aggregated-map segmentation methods into sparse-voxel production anchors, point-conv geometry baselines, superpoint graph map-scale context, serialized transformer/foundation-model ceilings, projection and standard-op deployment lanes, image-distilled/open-vocabulary label lanes, and SSM/Mamba experiments. |
| Bake-off controls | The method-library route now requires shared source-map acceptance packages, taxonomy, release-state labels, tile manifests, halo policy, optimizer budget, modality lanes, and semantic plus hygiene metrics before claiming one architecture beats another. |
Recurring LOSC Consolidation Promotion (2026-05-23)
Five read-only scouts checked LOSC source maturity, page schema, cross-links, count updates, and active queue priority. This bounded pass promoted LOSC as an atomic method page because the arXiv, OpenReview, Valeo project page, and official repository are public, while existing repo coverage only mentioned it inside overview/data-engine paragraphs.
| Cluster | Promoted or refreshed coverage |
|---|---|
| Open-vocabulary LiDAR pseudo-label consolidation | LOSC |
| Aggregated-map routing | Aggregated-map segmentation, open-vocabulary detection, data-engine, active-labeling, and annotation-tool pages now treat LOSC outputs as candidate or pseudo-labeled regions until reviewer, taxonomy, QA, manifest, and map-hygiene gates promote them. |
Recurring Map-Hygiene Workflow Refresh (2026-05-23)
Five read-only scouts checked the active aggregated-map segmentation queue and found that the map-hygiene ground-truth protocol had two useful but weakly connected owners: a canonical map-operations protocol and a safety-validation companion. This bounded pass kept both pages, made the ownership boundary explicit, and routed the protocol through semantic-map QA, publication gates, operational monitoring, regulatory evidence, README, INDEX, and the active loop queue.
| Cluster | Refreshed coverage |
|---|---|
| Map-hygiene protocol ownership | Airside Map Hygiene Ground Truth Protocol now owns layer labels, capture slices, reviewer dispositions, QA gates, and benchmark splits. |
| V&V workflow handoff | Airside Map Hygiene Ground Truth Protocol now acts as the benchmark and acceptance-output companion instead of an ambiguous duplicate. |
| Aggregated-map routing | Aggregated-map segmentation, map construction, publication gates, monitoring, and regulatory evidence now share the same static/movable/dynamic/FOD/artifact/unknown vocabulary. |
Recurring Aggregated-Map Selector Refresh (2026-05-23)
Five read-only scouts checked the compact selector candidate against existing dataset/proxy, training-paradigm, LOSC, UniLiPs, and map-hygiene coverage. This bounded pass kept the content in the aggregated-map hub instead of creating a new page, because the repo already has detailed companion pages for datasets, training paradigms, tiling, post-processing, and static-transient removal.
| Cluster | Refreshed coverage |
|---|---|
| Compact selector | Aggregated-Map Semantic Segmentation now includes a compact proxy/input/training selector for airside, dense urban districts, non-road sites, calibrated camera+LiDAR surveys, existing SLAM maps, long corridors, and mesh/digital-twin release. |
| Training companion routing | 3D Segmentation Training Paradigms now points readers to the hub selector for operational routing while keeping the detailed training trade-offs in one place. |
Recurring BeautyMap Dynamic-Removal Promotion (2026-05-23)
Five read-only scouts checked the aggregated-map segmentation and pointcloud-removal queue. This bounded pass promoted BeautyMap as the binary-encoded ground-matrix cleaner because it is official-code-backed, KTH-benchmark-routed, and directly supports the dynamic-removal conditioning stage before map-scale semantic segmentation.
Next perception promotion queue: LinkOcc remains a watchlist item until open method details or code make an atomic page reviewable. Other remaining items are e2e sparse-query perception follow-ons where they need separate pages, cooperative follow-ons such as CoDS, JigsawComm, GSCOOP-style collaborative Gaussian BEV segmentation, and airside-specific dust/de-icing/steam/glycol/wet-apron validation datasets. SparseDriveV2 and DiffusionDriveV2 are now routed through planning/E2E benchmark refreshes rather than new perception pages; DySS is routed through sparse-query and streaming-temporal overviews pending official code/adoption. DSERT-RoLL and CMHT are promoted as public multimodal adverse-weather dataset proxies, TruckV2X is promoted as a truck-centered V2X proxy, VOGS-CP is promoted as the collaborative Gaussian semantic-occupancy reference, EmbodiedScan/MMScan is promoted as the embodied 3D benchmark reference, SpaCeFormer is promoted as the proposal-free open-vocabulary 3D instance segmentation reference, LOSC is promoted as the open-vocabulary LiDAR pseudo-label consolidation reference, and Airport-FOD3S is promoted as the synthetic FOD data-engine reference, but no public dataset found in this loop directly covers airside dust, de-icing mist, steam, glycol film, or wet-apron multipath. Missing-view occupancy routes through M2-Occ, temporal radar-camera occupancy routes through TEOcc, Gaussian-rendered occupancy routes through GaussRender, and vision-only Gaussian occupancy label curation routes through GS-Occ3D unless a distinct method needs its own file.
P0 Discovery Clusters To Split Or Link
| Suggested file | Method or technique | Category | Why it matters | Primary sources |
|---|---|---|---|---|
camera-bev-depth-aware-encoders.md | BEVDet, BEVDepth, BEVStereo, BEVPoolv2, BEVDet4D, SOLOFusion | Depth-aware camera BEV | The current BEV doc jumps from LSS to BEVFormer; depth-supervised LSS-family encoders are more practical for LiDAR-supervised airside camera fallback. | https://arxiv.org/abs/2206.10092, https://arxiv.org/abs/2209.10248, https://arxiv.org/abs/2203.17054, https://openreview.net/forum?id=H3HcEJA2Um |
camera-only-3d-occupancy.md | TPVFormer, SurroundOcc, Occ3D/OpenOcc, FB-OCC, FlashOcc, SparseOcc, SelfOcc, RenderOcc | Camera semantic occupancy | Camera occupancy is now a core planning-facing alternative to boxes for arbitrary shapes, FOD, personnel, and unknown obstacles. | https://openaccess.thecvf.com/content/CVPR2023/html/Huang_Tri-Perspective_View_for_Vision-Based_3D_Semantic_Occupancy_Prediction_CVPR_2023_paper.html, https://openaccess.thecvf.com/content/ICCV2023/html/Wei_SurroundOcc_Multi-camera_3D_Occupancy_Prediction_for_Autonomous_Driving_ICCV_2023_paper.html, https://arxiv.org/abs/2304.14365, https://openaccess.thecvf.com/content/CVPR2024/html/Tang_SparseOcc_Rethinking_Sparse_Latent_Representation_for_Vision-Based_Semantic_Occupancy_Prediction_CVPR_2024_paper.html |
lidar-moving-object-segmentation.md | LiDAR-MOS, 4DMOS, InsMOS, MF-MOS, MV-MOS/CV-MOS, SegNet4D, Mask4D, HeLiMOS | Moving-object segmentation and 4D panoptic LiDAR | Airside needs point-level static/dynamic separation for parked-vs-moving GSE, map cleanup, track birth, and prediction gating. | https://arxiv.org/abs/2105.08971, https://arxiv.org/abs/2206.04129, https://arxiv.org/abs/2401.17023, https://arxiv.org/abs/2408.10602, https://arxiv.org/abs/2406.16279, https://github.com/url-kaist/HeLiMOS-PointCloud-Toolbox |
streaming-mos-4d-panoptic.md | StreamMOS and 4DSegStreamer | Streaming temporal segmentation | Adds memory across inferences, not only within one scan window; useful for high-FPS streams and occlusion consistency. | https://arxiv.org/abs/2407.17905, https://llada60.github.io/4DSegStreamer/ |
| Promoted: Instantaneous Motion Perception | Instantaneous perception of moving objects in 3D | Subtle motion estimation | Airside has creeping tugs and quasi-static objects; this fills the gap between single-frame detection and track-derived velocity. | https://openaccess.thecvf.com/content/CVPR2024/html/Liu_Instantaneous_Perception_of_Moving_Objects_in_3D_CVPR_2024_paper.html, https://arxiv.org/abs/2405.02781 |
4d-occupancy-forecasting.md | Cam4DOcc, StreamingFlow, UnO, DFIT-OccWorld, Drive-OccWorld | Perception-to-prediction handoff | Tracks alone cannot represent unknown, partially observed, or occluded movers; 4D occupancy/flow is a stronger planner interface. | https://openaccess.thecvf.com/content/CVPR2024/html/Ma_Cam4DOcc_Benchmark_for_Camera-Only_4D_Occupancy_Forecasting_in_Autonomous_Driving_CVPR_2024_paper.html, https://openaccess.thecvf.com/content/CVPR2024/html/Shi_StreamingFlow_Streaming_Occupancy_Forecasting_with_Asynchronous_Multi-modal_Data_Streams_via_CVPR_2024_paper.html, https://openaccess.thecvf.com/content/CVPR2024/html/Agro_UnO_Unsupervised_Occupancy_Fields_for_Perception_and_Forecasting_CVPR_2024_paper.html, https://arxiv.org/abs/2412.13772 |
4d-radar-perception-methods.md | K-Radar, View of Delft, RadarPillars, V2X-Radar | 4D radar perception | Radar is covered mainly as hardware/fusion; the perception stack needs tensor-vs-point representations, Doppler/RCS encoding, and benchmark limits. | https://github.com/kaist-avelab/K-Radar, https://intelligent-vehicles.org/datasets/view-of-delft/, https://arxiv.org/abs/2408.05020, https://arxiv.org/abs/2411.10962 |
multi-sensor-corruption-benchmarks.md | MSC-Bench and MultiCorrupt | Robustness and corruption evaluation | Adds systematic camera+LiDAR corruptions, missing inputs, misalignment, temporal mismatch, adverse weather, and sensor failures. | https://arxiv.org/abs/2501.01037, https://arxiv.org/abs/2402.11677, https://github.com/ika-rwth-aachen/MultiCorrupt |
real-sensor-anomaly-robustness.md | S2R-Bench and Occluded nuScenes | Real sensor anomaly and occlusion validation | Adds real adverse-weather/sensor-anomaly data and parameterized camera/radar/LiDAR occlusion tests beyond synthetic corruption overlays. | https://www.nature.com/articles/s41597-025-06255-3, https://arxiv.org/abs/2505.18631, https://arxiv.org/abs/2510.18552 |
open-world-ood-perception-benchmarks.md | OpenAD and S2M OOD object segmentation | Open-world/OOD evaluation | Current docs cover generic OOD scores; these add benchmarkable 3D corner cases and mask-level unknown-object localization. | https://arxiv.org/abs/2411.17761, https://github.com/VDIGPKU/OpenAD, https://openaccess.thecvf.com/content/CVPR2024/html/Zhao_Segment_Every_Out-of-Distribution_Object_CVPR_2024_paper.html |
open-world-open-vocab-detection.md | OW-OVD | Open-world plus open-vocabulary 2D detection | Existing open-vocab coverage emphasizes prompted known classes; OW-OVD adds unknown-object recall and incremental learning. | https://openaccess.thecvf.com/content/CVPR2025/html/Xi_OW-OVD_Unified_Open_World_and_Open_Vocabulary_Object_Detection_CVPR_2025_paper.html |
open-world-3d-objectness.md | OpenAD, OP3Det, WildDet3D, DetAny3D | Prompt-free and promptable 3D objectness | Fills the gap between closed-set 3D detection, text-prompted 3D detection, and true unknown-object discovery in 3D. | https://openreview.net/forum?id=T9UDyN5Tw6, https://arxiv.org/abs/2411.17761, https://openreview.net/forum?id=wEOmS8Aw1W, https://arxiv.org/abs/2510.17686, https://arxiv.org/abs/2604.08626, https://openaccess.thecvf.com/content/ICCV2025/papers/Zhang_Detect_Anything_3D_in_the_Wild_ICCV_2025_paper.pdf |
open-world-anomaly-segmentation-driving.md | Clipomaly, S2M, SegmentMeIfYouCan | Unknown-region and anomaly segmentation | Dedicated anomaly stack for OOD masks, RoadObstacle/RoadAnomaly metrics, and airside FOD adaptation. | https://arxiv.org/abs/2512.01427, https://arxiv.org/abs/2311.16516, https://segmentmeifyoucan.com/ |
sam3-concept-segmentation-airside.md | SAM 3 / SAM 3.1 concept segmentation and tracking | Foundation-model auto-labeling | Current docs stop at SAM/SAM2/Grounded-SAM2; SAM 3 adds open-vocabulary concept prompts, exemplars, and video tracking for GSE/FOD labeling. | https://arxiv.org/abs/2511.16719, https://github.com/facebookresearch/sam3 |
resilient-fusion-sensor-failure.md | MoME / nuScenes-R failure routing | Robust fusion under sensor failure | Directly handles LiDAR drop, camera drop, limited FOV, beam reduction, and occlusion as perception architecture, not only robustness theory. | https://openaccess.thecvf.com/content/CVPR2025/html/Park_Resilient_Sensor_Fusion_Under_Adverse_Sensor_Failures_via_Multi-Modal_Expert_CVPR_2025_paper.html, https://github.com/konyul/MoME |
robust-fusion-asynchrony-calibration.md | Staleness-aware fusion, GraphBEV, SOAC, RC-AutoCalib | Temporal sync and calibration drift | Covers silent fusion failure from stale sensors, extrinsic drift, LiDAR-camera misalignment, radar-camera calibration, and asynchronous inputs. | https://arxiv.org/abs/2506.05780, https://arxiv.org/abs/2403.11848, https://github.com/adept-thu/GraphBEV, https://openaccess.thecvf.com/content/CVPR2024/html/Herau_SOAC_Spatio-Temporal_Overlap-Aware_Multi-Sensor_Calibration_using_Neural_Radiance_Fields_CVPR_2024_paper.html, https://openaccess.thecvf.com/content/CVPR2025/html/Luu_RC-AutoCalib_An_End-to-End_Radar-Camera_Automatic_Calibration_Network_CVPR_2025_paper.html |
roadside-infrastructure-cooperative-datasets.md | RCooper and HoloVIC | Roadside/infrastructure cooperative perception datasets | Adds real multi-RSPU and multi-layout infrastructure perception, closer to airport fixed-sensor coverage than V2V-only datasets. | https://openaccess.thecvf.com/content/CVPR2024/html/Hao_RCooper_A_Real-world_Large-scale_Dataset_for_Roadside_Cooperative_Perception_CVPR_2024_paper.html, https://github.com/AIR-THU/DAIR-RCooper, https://openaccess.thecvf.com/content/CVPR2024/html/Ma_HoloVIC_Large-scale_Dataset_and_Benchmark_for_Multi-Sensor_Holographic_Intersection_and_CVPR_2024_paper.html, https://arxiv.org/abs/2403.02640 |
infrastructure-adverse-weather-cooperative-perception.md | CoInfra | Multi-node infrastructure plus adverse weather | Strong airport analogue: synchronized infrastructure nodes, real weather, 5G delay-aware fusion, OTA, monitoring, and system docs. | https://arxiv.org/abs/2507.02245, https://github.com/NingMingHao/CoInfra |
v2x-online-realtime-benchmarks.md | V2X-ReaLO | Online V2X framework and dataset | Existing docs are mostly offline-benchmark oriented; this targets synchronized ROS bags, latency, and online intermediate fusion. | https://arxiv.org/abs/2503.10034 |
cooperative-occupancy-perception.md | CoHFF collaborative semantic occupancy | Cooperative occupancy | Fills a major gap: dense 3D semantic occupancy, not only boxes, for irregular cargo, unknown obstacles, and FOD. | https://openaccess.thecvf.com/content/CVPR2024/html/Song_Collaborative_Semantic_Occupancy_Prediction_with_Hybrid_Feature_Fusion_in_Connected_CVPR_2024_paper.html, https://github.com/rruisong/CoHFF |
communication-efficient-cooperative-fusion.md | CoSDH | Bandwidth-efficient hybrid fusion | Supply-demand region selection plus intermediate/late hybridization is more deployment-oriented than pure intermediate fusion. | https://openaccess.thecvf.com/content/CVPR2025/html/Xu_CoSDH_Communication-Efficient_Collaborative_Perception_via_Supply-Demand_Awareness_and_Intermediate-Late_Hybridization_CVPR_2025_paper.html |
cooperative-sequential-tracking.md | CoopTrack | Cooperative MOT and sequential perception | Existing cooperative docs emphasize detection; airport planners need stable IDs and motion continuity across agents. | https://openaccess.thecvf.com/content/ICCV2025/html/Zhong_CoopTrack_Exploring_End-to-End_Learning_for_Efficient_Cooperative_Sequential_Perception_ICCV_2025_paper.html, https://arxiv.org/abs/2507.19239 |
uncertainty-detection-conformal-abstention.md | Conformal bounding boxes and conformal abstention | Detection-specific uncertainty | Upgrades conformal coverage from class sets to localization intervals and explicit abstain/degrade decisions. | https://arxiv.org/abs/2403.07263, https://arxiv.org/abs/2502.07255 |
| Promoted: AIDE | AIDE automatic data engine | Long-tail detection data engine | Concrete loop for issue mining, auto-labeling, scenario generation, and open-world detection evaluation. | https://openaccess.thecvf.com/content/CVPR2024/html/Liang_AIDE_An_Automatic_Data_Engine_for_Object_Detection_in_Autonomous_CVPR_2024_paper.html, https://arxiv.org/abs/2403.17373 |
30-autonomy-stack/perception/methods/vespa-open-world-pointcloud-labeling.md | VESPA and adjacent VLM-assisted 3D auto-labeling | Open-world point-cloud labeling | Multimodal pseudo-labeling without GT annotations or HD maps directly addresses airside 3D labeling bottlenecks; the ScienceDirect source is an adjacent VLM/pseudo-LiDAR auto-annotation reference, not a VESPA source. | https://arxiv.org/abs/2507.20397, https://www.sciencedirect.com/science/article/pii/S0968090X25004334 |
P1 Remaining Atomic Pages or Major Sections
| Suggested file | Method or technique | Category | Why it matters | Primary sources |
|---|---|---|---|---|
camera-panoptic-open-vocab-occupancy.md | PanoOcc, VEON, LangOcc, OpenOcc | Camera panoptic/open-vocabulary occupancy | Adds instance-level and open-vocabulary 3D scene understanding for rare GSE, barriers, FOD, and long-tail objects. | https://openaccess.thecvf.com/content/CVPR2024/html/Wang_PanoOcc_Unified_Occupancy_Representation_for_Camera-based_3D_Panoptic_Segmentation_CVPR_2024_paper.html, https://arxiv.org/abs/2407.12294, https://openreview.net/forum?id=KhjlXNbYea, https://github.com/boschresearch/LangOcc, https://arxiv.org/abs/2403.11796 |
camera-4d-occupancy-forecasting.md | Cam4DOcc, OccProphet, ALOcc | Temporal camera occupancy | Current temporal perception coverage is stronger on detection/tracking than camera 4D occupancy. | https://openaccess.thecvf.com/content/CVPR2024/html/Ma_Cam4DOcc_Benchmark_for_Camera-Only_4D_Occupancy_Forecasting_in_Autonomous_Driving_CVPR_2024_paper.html, https://openreview.net/forum?id=vC7AlY1ytz, https://huggingface.co/papers/2411.07725 |
| Promoted: Streaming Gaussian Occupancy | GaussianWorld and S2GO | Streaming Gaussian occupancy | Moves from per-frame occupancy toward persistent sparse Gaussian/query world state for real-time temporal occupancy. | https://openaccess.thecvf.com/content/CVPR2025/html/Zuo_GaussianWorld_Gaussian_World_Model_for_Streaming_3D_Occupancy_Prediction_CVPR_2025_paper.html, https://openreview.net/forum?id=z8ggdMlSco, https://arxiv.org/abs/2506.05473 |
generalized-uncalibrated-occupancy.md | OccAny and ViGT | Generalized / calibration-free occupancy | Targets uncalibrated, out-of-domain, monocular, sequential, and surround-view occupancy for heterogeneous camera rigs. | https://arxiv.org/abs/2603.23502, https://github.com/valeoai/OccAny, https://arxiv.org/abs/2602.05573 |
missing-view-resilient-occupancy.md | M2-Occ | Incomplete multi-camera robustness | Directly tests missing camera/view dropout for semantic occupancy. | https://arxiv.org/abs/2603.09737 |
occupancy-uncertainty-ood-evidence.md | ProOOD follow-on; EvOcc is promoted | Occupancy uncertainty / OOD | Handles unknown objects, occluded voxels, contradictory evidence, and long-tail classes at occupancy level without duplicating the existing EvOcc page. | https://openaccess.thecvf.com/content/CVPR2025/html/Kalble_EvOcc_Accurate_Semantic_Occupancy_for_Automated_Driving_Using_Evidence_Theory_CVPR_2025_paper.html, https://arxiv.org/abs/2604.01081 |
satellite-assisted-occupancy.md | SA-Occ | Satellite / overhead-map assisted occupancy | Especially relevant to airside because overhead imagery and GPS/IMU can reduce occlusion and distant-region weakness. | https://openaccess.thecvf.com/content/ICCV2025/html/Chen_SA-Occ_Satellite-Assisted_3D_Occupancy_Prediction_in_Real_World_ICCV_2025_paper.html |
| Promoted: Spatiotemporal Memory Occupancy Flow | ST-Occ and STCOcc | Temporal occupancy and scene flow | Adds memory, uncertainty-aware aggregation, sparse state refinement, and scene-flow prediction as a planner-facing interface. | https://openaccess.thecvf.com/content/ICCV2025/html/Leng_Occupancy_Learning_with_Spatiotemporal_Memory_ICCV_2025_paper.html, https://openaccess.thecvf.com/content/CVPR2025/html/Liao_STCOcc_Sparse_Spatial-Temporal_Cascade_Renovation_for_3D_Occupancy_and_Scene_CVPR_2025_paper.html |
| Promoted: Sparse Query Camera 3D Detection, SparseBEV, DETR4D, Sparse4D | Sparse4D, SparseBEV, DETR4D, object-query temporal memory | Sparse multi-view camera detection | Focused taxonomy and atomic method pages for sparse query methods versus dense BEV methods for edge deployment. | https://arxiv.org/abs/2211.10581, https://arxiv.org/abs/2308.09244, https://arxiv.org/abs/2212.07849, https://arxiv.org/abs/2602.00450 |
camera-depth-stereo-foundation-models.md | Depth Pro, Video Depth Anything, FoundationStereo | Camera depth and stereo foundation models | Updates fallback perception beyond Depth Anything v2, Metric3D, and UniDepth with metric, temporal, and zero-shot stereo options. | https://arxiv.org/abs/2410.02073, https://arxiv.org/abs/2501.12375, https://github.com/DepthAnything/Video-Depth-Anything, https://arxiv.org/abs/2501.09898, https://research.nvidia.com/labs/lpr/publication/stereoanything2025/ |
camera-lidar-fusion-interfaces.md | FUTR3D, CMT, DeepInteraction, AutoAlignV2, DAL, SAMFusion, MS-Occ | Camera-LiDAR fusion interfaces | Existing docs cover BEVFusion but underrepresent query fusion, calibration-tolerant alignment, dynamic modality dropout, and occupancy fusion. | https://arxiv.org/abs/2203.10642, https://arxiv.org/abs/2301.01283, https://proceedings.neurips.cc/paper_files/paper/2022/hash/0d18ab3b5fabfa6fe47c62e711af02f0-Abstract-Conference.html, https://arxiv.org/abs/2207.10316, https://arxiv.org/abs/2311.07152, https://arxiv.org/abs/2508.16408, https://arxiv.org/abs/2504.15888 |
camera-bev-robustness-validation.md | RoboBEV, BEV corruptions, viewpoint robustness, missing camera inputs, occupancy uncertainty | Camera BEV validation | Airside camera rigs, soiling, glare, missing cameras, and calibration changes need benchmark coverage. | https://arxiv.org/abs/2304.06719, https://arxiv.org/abs/2405.17426, https://arxiv.org/abs/2309.05192, https://arxiv.org/abs/2406.11021, https://arxiv.org/abs/2603.09737 |
adverse-weather-radar-lidar-3d-detection.md | LiDAR-4D radar fusion, L4DR, V2X-R | Radar-LiDAR adverse-weather detection | Directly addresses when radar should rescue LiDAR detection under fog/rain. | https://openaccess.thecvf.com/content/CVPR2024/papers/Chae_Towards_Robust_3D_Object_Detection_with_LiDAR_and_4D_Radar_CVPR_2024_paper.pdf, https://arxiv.org/abs/2408.03677, https://openaccess.thecvf.com/content/CVPR2025/papers/Huang_V2X-R_Cooperative_LiDAR-4D_Radar_Fusion_with_Denoising_Diffusion_for_3D_CVPR_2025_paper.pdf |
| Promoted: DepthOcc | DepthOcc | Depth-aware camera occupancy | Real-time camera semantic occupancy using monocular/stereo depth fusion and temporal enhancement; keep separate from radar-camera queues unless a method adds radar inputs. | https://doi.org/10.1016/j.cviu.2026.104670, https://tsinghua-mars-lab.github.io/Occ3D/ |
| LinkOcc routing/watchlist | LinkOcc temporal-association occupancy plus separate radar-camera follow-ons | Temporal occupancy and source hygiene | LinkOcc is IEEE TCSVT 2025 "3D Semantic Occupancy Prediction With Temporal Association"; do not cite RobuRCDet's arXiv ID as LinkOcc or classify LinkOcc as radar-camera without primary-source evidence. | https://ieeexplore.ieee.org/document/10734407, https://dblp.org/rec/journals/tcsv/OuyangXSWX25.html, https://arxiv.org/abs/2502.13071 |
| Promoted: 4D Radar-Camera Occupancy | 4DRC-OCC and RadarOcc | Radar-camera semantic occupancy | Dense occupancy using 4D radar plus camera is distinct from LiDAR-camera occupancy and generic dynamic occupancy; keep this row as provenance and watch radar-occupancy follow-ons. | https://arxiv.org/abs/2603.07794 |
| Promoted: Availability-Aware Sensor Fusion | Availability-aware Sensor Fusion via Unified Canonical Space | Sensor-failure perception | Directly targets missing/degraded camera, LiDAR, and 4D radar availability instead of assuming all modalities are valid. | https://arxiv.org/abs/2503.07029 |
event-camera-3d-detection.md | Ev-3DOD / DSEC-3DOD and eAP | Event-camera 3D perception | Event cameras now support 3D detection and time-to-collision, not just 2D detection or VIO. | https://arxiv.org/abs/2502.19630, https://openaccess.thecvf.com/content/CVPR2025/papers/Cho_Ev-3DOD_Pushing_the_Temporal_Boundaries_of_3D_Object_Detection_with_CVPR_2025_paper.pdf, https://arxiv.org/abs/2603.16303 |
fmcw-lidar-perception.md | AevaScenes and POD predictive object detection | FMCW / 4D LiDAR perception | Per-point velocity changes detection, scene-flow, and predictive-detection design compared with pulsed LiDAR. | https://www.aeva.com/press/aeva-introduces-aevascenes-the-first-open-access-fmcw-4d-lidar-and-camera-dataset-for-autonomous-vehicle-research/, https://arxiv.org/abs/2504.05649 |
| Promoted: 4D Radar Road Boundaries and Freespace | 4DRadarRBD | Radar freespace and road-boundary detection | Radar-based boundary/freespace perception is separate from object detection and occupancy, especially in poor visibility. | https://arxiv.org/abs/2503.01930, https://www.frontiersin.org/articles/10.3389/frsip.2025.1667789/ |
fully-sparse-lidar-detectors.md | VoxelNeXt, DSVT, fully sparse detector upgrade path | LiDAR 3D detection | openpcdet-centerpoint.md is CenterPoint-heavy; modern candidates should include sparse detectors with better dense-BEV cost profiles. | https://github.com/open-mmlab/OpenPCDet, https://arxiv.org/abs/2303.11301, https://arxiv.org/abs/2301.06051 |
thermal-lidar-dense-perception.md | Thermal-LiDAR depth completion, RGB-T-LiDAR segmentation, dense night freespace | Thermal/night dense perception | Current thermal doc is detection/safety-use-case oriented; dense depth and segmentation matter for freespace and occupancy. | https://arxiv.org/abs/2504.02356, https://doi.org/10.1016/j.isprsjprs.2026.01.008, https://oem.flir.com/en-gb/solutions/automotive/adas-dataset-form/, https://soonminhwang.github.io/rgbt-ped-detection/, https://github.com/bupt-ai-cz/LLVIP |
| Promoted: Dynamic Occupancy and Freespace | DIO, MS-Occ, UniOcc | Current-state occupancy and freespace | Perception docs need the current-state safety view that planners consume, distinct from predictive world models. | https://openaccess.thecvf.com/content/CVPR2025/papers/Diehl_DIO_Decomposable_Implicit_4D_Occupancy-Flow_World_Model_CVPR_2025_paper.pdf, https://arxiv.org/abs/2504.15888, https://arxiv.org/abs/2503.24381 |
| Promoted: LiDAR-Camera Occupancy Fusion | SDG-OCC and MS-Occ | LiDAR-camera occupancy fusion | Occupancy-specific fusion uses LiDAR geometry plus camera semantics through depth guidance, distillation, and multi-stage fusion. | https://openaccess.thecvf.com/content/CVPR2025/html/Duan_SDGOCC_Semantic_and_Depth-Guided_Birds-Eye_View_Transformation_for_3D_Multimodal_CVPR_2025_paper.html, https://arxiv.org/abs/2504.15888 |
uncertainty-aware-bev-fusion.md | GaussianLSS and HyperDUM | Fusion-specific uncertainty | Adds BEV depth uncertainty and low-cost feature-level multimodal uncertainty closer to deployed fusion stacks. | https://openaccess.thecvf.com/content/CVPR2025/html/Lu_Toward_Real-world_BEV_Perception_Depth_Uncertainty_Estimation_via_Gaussian_Splatting_CVPR_2025_paper.html, https://openaccess.thecvf.com/content/CVPR2025/html/Chen_Hyperdimensional_Uncertainty_Quantification_for_Multimodal_Uncertainty_Fusion_in_Autonomous_Vehicles_CVPR_2025_paper.html |
night-thermal-event-radar-benchmarks.md | NSAVP and V2X-Radar | Night, thermal, event, radar evaluation | Updates night perception beyond FLIR/KAIST-style references with stereo thermal, event cameras, 4D radar, rain/night/dusk cooperative data. | https://umautobots.github.io/nsavp, https://arxiv.org/abs/2401.13853, https://arxiv.org/abs/2411.10962 |
edge-real-time-bev-scheduling.md | RT-BEV and Jigsaw | Deadline-aware BEV deployment | Compression doc is TensorRT-heavy; add P99/WCET scheduling and dual-SoC partitioning. | https://rtcl.eecs.umich.edu/rtclweb/assets/publications/2024/rtss24-liu.pdf, https://2024.rtss.org/accepted-papers/index.html, https://dblp.org/rec/conf/rtss/SunLHHX0BSRG24.html |
40-runtime-systems/ros-autoware/sensor-in-memory-ros2-av-middleware.md | SIM shared-memory transport for ROS 2 / Autoware | Deployment latency and middleware | Targets perception-to-decision tail latency, not only model inference time. | https://arxiv.org/abs/2510.11448 |
ood-data-engine.md | DriveGEN and active OOD generation | Active data and evaluation | Complements uncertainty-triggered logging with controllable OOD augmentation and scenario generation. | https://openaccess.thecvf.com/content/CVPR2025/html/Lin_DriveGEN_Generalized_and_Robust_3D_Detection_in_Driving_via_Controllable_CVPR_2025_paper.html |
uncertainty-aware-end-to-end-3d-mot.md | S2-Track | Tracking uncertainty | Query initialization, probabilistic decoder, and uncertainty as downstream prediction signal are currently underdocumented. | https://proceedings.mlr.press/v267/tang25p.html, https://openreview.net/forum?id=vHr9cdeFfu, https://arxiv.org/abs/2406.02147 |
relation-aware-3d-mot.md | GRAE-3DMOT | Relation-aware tracking | Complements MCTrack/3DMOTFormer with relation-aware association for dense scenes. | https://openaccess.thecvf.com/content/CVPR2025/html/Kim_GRAE-3DMOT_Geometry_Relation-Aware_Encoder_for_Online_3D_Multi-Object_Tracking_CVPR_2025_paper.html |
latency-aware-streaming-perception.md | LASP online benchmark, Transtreaming, multi-modal streaming 3D detection | Latency compensation and online evaluation | Separate acquisition latency, inference latency, output timestamp compensation, and time-stale quality on edge hardware. | https://arxiv.org/abs/2504.19115, https://arxiv.org/abs/2409.06584, https://arxiv.org/abs/2209.04966 |
4d-world-model-pretraining.md | DriveWorld | Temporal representation pretraining | Concrete 4D pretraining bridge across detection, tracking, forecasting, occupancy, and planning. | https://arxiv.org/abs/2405.04390, https://openaccess.thecvf.com/content/CVPR2024/html/Min_DriveWorld_4D_Pre-trained_Scene_Understanding_via_World_Models_for_Autonomous_CVPR_2024_paper.html |
object-query-cooperative-perception.md | CoopDETR | Sparse cooperative communication | Object-query sharing is interpretable and extremely bandwidth-light versus region/BEV maps. | https://arxiv.org/abs/2502.19313 |
collaboration-robust-cooperative-fusion.md | mmCooper | Robust multi-stage fusion | Targets bandwidth limits plus calibration/misalignment errors in cooperative perception. | https://openaccess.thecvf.com/content/ICCV2025/html/Liu_mmCooper_A_Multi-agent_Multi-stage_Communication-efficient_and_Collaboration-robust_Cooperative_Perception_Framework_ICCV_2025_paper.html |
spatiotemporal-cooperative-perception.md | CoST | Space-time cooperative fusion | Unifies multi-agent and temporal fusion; avoids retransmitting static-object features repeatedly. | https://openaccess.thecvf.com/content/ICCV2025/html/Tang_CoST_Efficient_Collaborative_Perception_From_Unified_Spatiotemporal_Perspective_ICCV_2025_paper.html |
end-to-end-v2x-cooperative-driving.md | V2Xverse / CoDriving, UniV2X, Select2Drive | Closed-loop V2X perception-to-planning | Moves evaluation from mAP-only to driving score, collision rate, latency, and communication-aware planning. | https://collaborativeperception.github.io/V2Xverse/, https://github.com/CollaborativePerception/V2Xverse, https://arxiv.org/abs/2404.00717 |
privacy-adaptation-v2x-collaboration.md | Unknown-collaborator privacy and CoPEFT | Privacy/domain adaptation | Covers separately trained agents, privacy-preserving fusion, and low-cost per-site adaptation for multi-vendor fleets. | https://mlanthology.org/aaai/2025/lu2025aaai-privacy/, https://ojs.aaai.org/index.php/AAAI/article/view/34502 |
auto-vocabulary-lidar-segmentation.md | 3D-AVS | LiDAR open-vocabulary segmentation | Distinguishes LiDAR foundation pretraining from runtime vocabulary discovery for point clouds. | https://openaccess.thecvf.com/content/CVPR2025/html/Wei_3D-AVS_LiDAR-based_3D_Auto-Vocabulary_Segmentation_CVPR_2025_paper.html, https://arxiv.org/abs/2406.09126 |
mosaic3d-open-vocab-3d-segmentation.md | Mosaic3D | Open-vocabulary 3D segmentation | Foundation dataset/model for open-vocabulary 3D semantic and instance segmentation; fills a gap beyond 2D open-vocabulary segmentation. | https://research.nvidia.com/labs/lpr/publication/choy2025mosaic/, https://openaccess.thecvf.com/content/CVPR2025/papers/Lee_Mosaic3D_Foundation_Dataset_and_Model_for_Open-Vocabulary_3D_Segmentation_CVPR_2025_paper.pdf |
openvox-open-vocab-voxel-mapping.md | OpenVox | Open-vocabulary voxel mapping | Real-time instance-level probabilistic voxel memory is relevant for persistent apron maps and language-queryable scene memory. | https://open-vox.github.io/ |
ago-open-world-3d-occupancy.md | AGO adaptive grounding | Open-world 3D occupancy | Known/unknown semantic occupancy using VLM grounding is a stronger safety-layer fit than box-only open-world detection. | https://arxiv.org/abs/2504.10117, https://openaccess.thecvf.com/content/ICCV2025/papers/Li_AGO_Adaptive_Grounding_for_Open_World_3D_Occupancy_Prediction_ICCV_2025_paper.pdf |
| Promoted: Open3DTrack | Open3DTrack | Open-vocabulary 3D tracking | First-pass gaps covered 3D detection/objectness but not persistent open-vocabulary 3D tracks. | https://arxiv.org/abs/2410.01678 |
open-vocab-monocular-3d-detection.md | OVM3D-Det | Open-vocab monocular 3D detection | Useful as offline/assistive camera fallback, with safety caveats. | https://ovm3d-det.github.io/, https://arxiv.org/abs/2411.15657 |
dense-open-vocab-segmentation.md | Florence-2, CAT-Seg, ProxyCLIP, AutoSeg/AVS | Dense open-vocabulary labels | Grounded-SAM is not the whole dense-labeling story; pixel-level OV segmentation helps mask QA and rare-class discovery. | https://arxiv.org/abs/2311.06242, https://huggingface.co/microsoft/Florence-2-large, https://openaccess.thecvf.com/content/CVPR2024/html/Cho_CAT-Seg_Cost_Aggregation_for_Open-Vocabulary_Semantic_Segmentation_CVPR_2024_paper.html, https://arxiv.org/abs/2408.04883, https://openaccess.thecvf.com/content/ICCV2025/html/Ulger_Auto-Vocabulary_Semantic_Segmentation_ICCV_2025_paper.html |
ovad-open-vocab-3d-attributes.md | OVAD / OVODA | Open-vocabulary 3D attributes | Adds attribute-level open-vocabulary 3D detection, which matters for airside state labels such as cones, tugs, carts, lights, covers, and unusual vehicle configurations. | https://arxiv.org/abs/2508.16812 |
| Covered: VLM/VLA Reliability Benchmarks | DriveBench | VLM driving reliability under corruptions | Tests clean, corrupted, and text-only driving QA to catch visually ungrounded VLM perception or monitoring outputs. Keep as a cross-link, not a duplicate method page. | https://drive-bench.github.io/, https://github.com/worldbench/DriveBench |
airside-fod-open-world-benchmarks.md | FOD-A, FAA FOD framing, Airport-FOD3S, RDD5000, DualFOD/FOD-UAS | Residual airside FOD OOD benchmark plan | FOD-A, Airport-FOD3S, RDD5000, and DualFOD/FOD-UAS are now routed or promoted; the residual gap is a dedicated benchmark plan for tiny-object recall, OOD-vs-detector comparison, and public airside dust/de-icing/steam/glycol/wet-apron negatives. | https://arxiv.org/abs/2110.03072, https://github.com/FOD-UNOmaha/FOD-data, https://www.faa.gov/airports/airport_safety/fod, https://www.mdpi.com/1424-8220/25/15/4565, https://www.mdpi.com/2072-4292/17/4/669, https://www.mdpi.com/2504-446X/10/3/225 |
| Promoted: Airport-FOD3S Synthetic FOD Data Engine | Airport-FOD3S | Synthetic FOD data engine | Synthetic FOD generation and blending directly addresses rare small-object data scarcity beyond FOD-A, with data-lineage and real-only validation gates. | https://www.mdpi.com/1424-8220/25/15/4565 |
| Promoted: Fail2Drive | Fail2Drive | Closed-loop OOD validation | Paired-route distribution-shift scenarios expose perception-linked failures under closed-loop drift. | https://arxiv.org/abs/2604.08535, https://github.com/autonomousvision/fail2drive |
P2 or Watchlist
| Method or technique | Why watch | Current concern | Sources |
|---|---|---|---|
| VLMFusionOcc3D, OccVLA, O3N | VLM-assisted and open-vocabulary occupancy are moving quickly. | Promising but not yet safety perception primitives; validate hallucination, latency, and reproducibility. | https://arxiv.org/abs/2603.02609, https://huggingface.co/papers/2509.05578, https://arxiv.org/abs/2603.12144 |
| RayD3D, Dr.Occ, SuperOcc, Gau-Occ | Useful 2025-2026 camera/multimodal/Gaussian occupancy signals. | Mostly preprint-stage or narrow-benchmark. GaussRender, GaussianFlowOcc, GaussTR, and GS-Occ3D are promoted as method pages. | https://huggingface.co/papers/2603.22852, https://arxiv.org/abs/2603.01007, https://openaccess.thecvf.com/content/CVPR2025/html/Oh_3D_Occupancy_Prediction_with_Low-Resolution_Queries_via_Prototype-aware_View_Transformation_CVPR_2025_paper.html, https://gs-occ3d.github.io/ |
| EventFly, EvDET200K, TUMTraf EMOT, PRE-Mamba | Event perception is expanding into dense segmentation, high-definition detection, MOT, and rain removal. | Useful follow-ons for event-camera pages; promote after dataset/code adoption and hardware fit are clearer. | https://openaccess.thecvf.com/content/CVPR2025/html/Kong_EventFly_Event_Camera_Perception_from_Ground_to_the_Sky_CVPR_2025_paper.html, https://openaccess.thecvf.com/content/CVPR2025/html/Wang_Object_Detection_using_Event_Camera_A_MoE_Heat_Conduction_based_CVPR_2025_paper.html, https://arxiv.org/abs/2512.14595, https://arxiv.org/abs/2505.05307 |
| DSERT-RoLL | Stereo event/RGB/thermal, 4D radar, and dual-LiDAR dataset for diverse driving conditions. | Promoted into DSERT-RoLL; keep watching for benchmark/code updates and license changes. | https://arxiv.org/abs/2604.03685 |
| HyperDet, DRIFT, SD4R, R4Det | 2026 radar-only/radar-camera perception line. | Needs code, repeatable benchmarks, and multi-dataset validation. | https://arxiv.org/abs/2602.11554, https://arxiv.org/abs/2603.09695, https://arxiv.org/abs/2602.20653, https://arxiv.org/abs/2603.11566 |
| SiMO and single-modality-operable cooperative perception | Directly relevant to robust collaborative perception under sensor failures. | Too fresh; compare against MoME, HEAL, mmCooper, and CoSDH before promotion. | https://gist.science/paper/2603.08240 |
| CoDS, JigsawComm, and GSCOOP | 2025-2026 cooperative perception wave around semantic communication, compression, utility scheduling, and sparse Gaussian sharing. | Useful but not yet mature enough for core docs; CoDS still lacks code/project/final-venue artifacts, JigsawComm has only a placeholder repository, and GSCOOP's official repo link is currently 404/private despite public ICCV paper/poster sources. Compare against CoSDH, mmCooper, CoST, V2X-ReaLO, promoted VOGS-CP, promoted TruckV2X, promoted SparseCoop, and promoted QuantV2X first. | https://arxiv.org/abs/2512.22513, https://arxiv.org/abs/2511.17843, https://github.com/WiSeR-Lab/JigsawComm, https://openaccess.thecvf.com/content/ICCV2025/html/Hong_Communication-Efficient_Multi-Vehicle_Collaborative_Semantic_Segmentation_via_Sparse_3D_Gaussian_Sharing_ICCV_2025_paper.html |
| LightDiff and JarvisIR | Restoration can help low-light/weather preprocessing. | Treat as monitored advisory preprocessing, not safety authority, because restoration can hallucinate. | https://openaccess.thecvf.com/content/CVPR2024/html/Li_Light_the_Night_A_Multi-Condition_Diffusion_Framework_for_Unpaired_Low-Light_CVPR_2024_paper.html, https://openaccess.thecvf.com/content/CVPR2025/html/Lin_JarvisIR_Elevating_Autonomous_Driving_Perception_with_Intelligent_Image_Restoration_CVPR_2025_paper.html |
| T-Rex2 and OWLv2 visual-prompt detection | Visual exemplars help airport-specific GSE variants where text prompts are brittle. | Better for onboarding and annotation acceleration than runtime safety perception. | https://arxiv.org/abs/2403.14610, https://github.com/IDEA-Research/T-Rex, https://arxiv.org/abs/2306.09683 |
| DySS, GaussianDet3D, OffsetOcc, Diffusion-FS | Useful follow-ons for streaming camera detection, Gaussian camera detection, panoptic completion, and corridor prediction. | DySS is routed through existing sparse-query and streaming-temporal overviews until official code or broader adoption justifies an atomic page. The others remain useful research directions, but merge into existing occupancy, SAM/data-engine, or freespace pages unless adoption grows. ForeSight and SAM4D already have atomic pages. | https://arxiv.org/abs/2506.10242, https://openaccess.thecvf.com/content/CVPR2025W/WAD/html/Yasarla_DySS_Dynamic_Queries_and_State-Space_Learning_for_Efficient_3D_Object_CVPRW_2025_paper.html, https://openreview.net/forum?id=qByiOX1j9C, https://cvpr.thecvf.com/virtual/2025/35771, https://arxiv.org/abs/2507.18763 |
Benchmark and Dataset Gaps
Already Covered, But Needs Better Discoverability
| Existing file | Add aliases or cross-links for |
|---|---|
| BEV Encoding Architectures | BEVDet, BEVDepth, BEVStereo, BEVPoolv2, BEVDet4D, SOLOFusion, DuoSpaceNet, DenseBEV, SEPatch3D. |
| Camera-Only Degraded Perception | Depth Pro, Video Depth Anything, FoundationStereo, OVM3D-Det, camera occupancy fallback, MultiCorrupt/MSC-Bench degradation tests. |
| Open-Vocabulary and Zero-Shot Detection | Split open-vocabulary known-class detection from open-world unknown-object discovery; add OW-OVD, SAM 3, OP3Det, 3D-AVS, Clipomaly, S2M, T-Rex2. |
| Vision Foundation Models | Florence-2, CAT-Seg, ProxyCLIP, AutoSeg/AVS, SAM 3, data-engine use cases, and OOD segmentation. |
| LiDAR Foundation Models | 3D-AVS, OpenAD, OP3Det, open-world objectness, and auto-vocabulary point-cloud segmentation. |
| LiDAR Semantic Segmentation | LiDAR-MOS, 4DMOS, MotionSeg3D, MambaMOS, SegNet4D, Mask4D, ALPINE, HeLiMOS, moving/static labels, scene-flow signals, and 4D panoptic outputs. |
| OpenPCDet and CenterPoint | VoxelNeXt, DSVT, fully sparse detectors, RadarPillars, and camera-LiDAR/radar fusion interfaces. |
| Streaming Temporal Perception | StreamMOS, 4DSegStreamer, MotionSeg3D, MambaMOS, Neural Scene Flow Priors, Cam4DOcc, UnO, DFIT-OccWorld, Drive-OccWorld, LASP, Transtreaming, acquisition-vs-inference latency. |
| Multi-Object Tracking | S2-Track, GRAE-3DMOT, CoopTrack, uncertainty/existence output contracts, occlusion age, source modality, track history quality. |
| Infrastructure Cooperative Perception | RCooper, HoloVIC, CoInfra, V2X-ReaLO, QuantV2X, TruckV2X, VOGS-CP, V2XScenes, UrbanIng-V2X, V2X-Radar, online realism and latency traces. |
| Collaborative Fleet Perception | CoSDH, SparseCoop, QuantV2X, VOGS-CP, mmCooper, CoST, CoopDETR, CoopTrack, CoHFF, V2Xverse, privacy/adaptation for unknown collaborators. |
| Uncertainty Quantification | Conformal boxes, conformal abstention, GaussianLSS, HyperDUM, segmentation/occupancy/freespace uncertainty, modality health. |
| Model Compression and Edge Deployment | QuantV2X, RT-BEV, Jigsaw, P99/WCET gates, multi-model contention, compression-vs-robustness regression checks. |
| Night Operations and Thermal Fusion | NSAVP, V2X-Radar, thermal-LiDAR dense perception, event cameras, thermal calibration drift, hot-engine clutter, rain-on-lens. |
| Production Perception Systems | Production validation matrix: missing modality, stale modality, miscalibration, sensor occlusion, darkness, fog/rain, degraded compute. |
| Data Flywheel, Data Engines, and Airport-FOD3S Synthetic FOD Data Engine | AIDE, SAM 3, Florence-2, OOD triggers, cooperative metadata, FOD-specific QA gates, synthetic sample lineage, and real-only holdout rules. |
Guardrail Process
When adding or revising perception research:
- Check whether the method is already a dedicated file, an existing section, or a backlog item in this audit.
- If it is P0, create one atomic file under
30-autonomy-stack/perception/methods/before expanding lower-priority coverage. - If it is P1, either create one atomic method file or add aliases and cross-links from the closest existing method page.
- If it is P2/watchlist, add aliases only after primary sources, code, or dataset availability are confirmed.
- Keep family synthesis in top-level perception docs and detailed method evidence in
methods/. - After every perception expansion, update this audit, Research Index, README, and any related safety/data-engine docs.