MOE LiDAR Moving Event Benchmark

Last updated: 2026-05-09

Why It Matters

MOE is a dense LiDAR Moving Event Detection (MED) dataset and benchmark. It targets a narrower question than semantic segmentation: which LiDAR points are generated by moving objects or moving events?

That distinction matters for static map building, dynamic obstacle filtering, and airside map QA. A moving aircraft tug, pedestrian, bus, or baggage cart can leave ghost structure in a LiDAR map even if the object class is known.

Dataset Snapshot

Field	MOE detail	Practical note
Full name	MOE: A Dense LiDAR MOving Event Dataset, Detection Benchmark and LeaderBoard	Treat as a MED benchmark, not a general semantic segmentation dataset
Source types	Ten sequences from high-fidelity simulation and real-world / campus-style scenes	Useful density and diversity, but still not an airport apron dataset
Public structure	Sequence folders such as 00, 01, 02; each contains pcd, label, and gt_poses.txt where labels are released	Supports reproducible replay with ground-truth poses
Point cloud format	xxxxxx.pcd with x,y,z information	Simple to load with Open3D or PCL
Label row	Moveable_ID Moving_Status Class_ID	Separates movable object membership from moving status
Pose format	KITTI-style first three rows of a 4x4 transform in gt_poses.txt	Convenient for MOS and map-cleaning baselines
Competition split	Sequences 05-09 are used for the CodaLab MED competition	Those sequences publish point clouds and poses without public motion labels

Benchmark Table

The MOE repository reports mean IoU over sequences 00, 01, and 02, with moving points treated as positive. The table below is the repository benchmark, not a live leaderboard scrape.

Method	Family	Seq 00 IoU	Seq 01 IoU	Seq 02 IoU	Mean IoU	Transfer note
DOD	Online non-learning	0.786	0.142	0.595	0.508	Strong in dense indoor/crowd-style cases; weaker on outdoor sequence 01
InsMOS	Learning-based MOS	0.495	0.282	0.379	0.385	Best learning baseline in the published MOE table, but trained on SemanticKITTI
ERASOR	Offline non-learning	0.378	0.028	0.627	0.344	Map-cleaning baseline; not designed for low-latency event output
Octomap	Offline non-learning	0.328	0.031	0.652	0.337	Occupancy baseline; sensitive to scene and range settings
Dynablox	Online non-learning	0.320	0.195	0.492	0.336	Incremental mapping flavor; useful for runtime dynamic masking
Removert	Offline non-learning	0.297	0.028	0.421	0.249	Good static-map reference but weak on MOE sequence 01
M-detector	Online non-learning MED	0.305	0.174	0.044	0.174	Point-stream detector; repository score reflects this dataset/protocol, not its latency advantage
MotionBEV	Learning-based BEV	0.002	0.055	0.069	0.042	Shows poor cross-domain transfer when trained elsewhere

Metrics

Metric	Definition / reporting guidance	Why it matters
Moving-event IoU	TP / (TP + FP + FN) for moving points	Core MOE ranking metric
Per-sequence IoU	Report each sequence, not only mean IoU	MOE results vary sharply by scene structure
Mean of per-scan IoU	Average IoU scan by scan when required by the protocol	More sensitive to sparse dynamic frames than pooled IoU
Dynamic precision	Fraction of predicted moving points that are truly moving	Controls false dynamic erosion of static map structure
Dynamic recall	Fraction of true moving points detected	Controls ghost leakage into maps and occupancy history
Latency	Point-level, scan-level, and P95/P99 processing delay	MED is valuable only if the result arrives before mapping/planning consumes the points
Pose sensitivity	Score under pose noise, timestamp drift, and extrinsic perturbation	MOE provides poses, but real fleets do not get perfect alignment

Airside Transfer

MOE asset	Use it for	Do not claim until local data exists
Dense moving-object scenes	Stress moving/static separation under crowded conditions	Aircraft pushback, towbar coupling, belt-loader alignment, and GSE staging
Simulation sequences	Verify label parsing, pose format, and metric code	Domain performance on reflective apron concrete and airport lighting
Public baselines	Compare offline cleaners, online occupancy methods, and learned MOS under one protocol	A single production default; the published table has strong sequence dependence
CodaLab held-out split	External regression signal if submitting methods	Safety-case evidence; leaderboard scores are not airside acceptance tests

Validation Guidance

1. Reproduce the repository benchmark on sequences 00-02 before adding private data.
2. Keep sequence-level scores in dashboards; MOE sequence 01 behaves very differently from 00 and 02.
3. Evaluate both online point output and delayed frame output when a method supports both.
4. Add static-preservation metrics alongside moving IoU if the output feeds map cleaning.
5. For airside replay, label movable-static states separately: parked GSE, staged carts, chocks, cones, stopped aircraft, and starting-to-move actors.
6. Use MOE to choose candidate algorithms, then retune thresholds on local sensor range, scan pattern, vehicle speed, and localization quality.

Sources

• MOE project site: https://sites.google.com/view/moe-dataset
• MOE dataset repository: https://github.com/DeepDuke/MOE-Dataset
• MOE CodaLab competition: https://codalab.lisn.upsaclay.fr/competitions/18028
• MOE IROS 2024 record: https://researchportal.hkust.edu.hk/en/publications/moe-a-dense-lidar-moving-event-dataset-detection-benchmark-and-le-2/
• M-detector paper: https://www.nature.com/articles/s41467-023-44554-8
• Local context: 30-autonomy-stack/perception/datasets-benchmarks/moving-static-separation-mos-datasets.md