EvOcc

What It Is

EvOcc is a CVPR 2025 framework for evidential semantic occupancy mapping and prediction.
The full title is "EvOcc: Accurate Semantic Occupancy for Automated Driving Using Evidence Theory."
It has two linked pieces: evidential reconstruction of semantic occupancy ground truth from noisy LiDAR and training of camera models to predict evidential occupancy.
It uses evidence theory to represent support, conflict, and unknown state rather than forcing every voxel into one hard class.
It is an uncertainty-aware occupancy supervision method and prediction target.
It is not only an inference-time calibration wrapper; it changes how the ground truth occupancy map is built.

Extend evidential mapping from binary occupancy to semantic occupancy.
Represent key hypotheses such as free, occupied-by-class, generally occupied with unknown semantics, and total uncertainty.
Build a semantic basic belief assignment from LiDAR transmissions, labeled reflections, and unlabeled reflections.
Explicitly model conflicting evidence, such as free-space rays crossing areas with noisy returns.
Use the evidential occupancy maps as ground truth to train multi-camera semantic occupancy predictors.
Predict both semantic occupancy and uncertainty so planners can see where the model is unsure due to occlusion or conflicting measurements.

Ground-truth construction input: LiDAR scans, semantic point labels where available, ray transmissions, and calibration.
Model inference input: multi-view camera images with camera calibration and ego-pose metadata.
Training target: semantic occupancy belief masses and uncertainty derived from evidence theory.
Output: semantic occupancy voxel grid.
Output: voxel-level uncertainty or belief mass over uncertainty-related hypotheses.
Optional output: BEV uncertainty map for planner gating and data mining.

Convert LiDAR scans into transmission and reflection evidence per voxel.
Assign support to semantic class hypotheses when labeled reflections are available.
Assign support to a general occupied hypothesis when the voxel is occupied but the semantic class is unknown.
Track free-space evidence from ray transmissions.
Combine supporting and conflicting evidence into belief masses for the selected hypothesis set.
Train a camera-based occupancy model with an evidential loss that matches both semantics and uncertainty.
Use the predicted uncertainty map downstream instead of discarding it after semantic argmax.

EvOcc is evaluated on semantic occupancy prediction for automated driving.
The paper reports that replacing heuristic LiDAR occupancy labels with evidential mapping substantially improves reconstructed occupancy quality.
It reports improved training outcomes for camera occupancy models while keeping the model architecture fixed and changing the supervision.
The CVPR paper emphasizes uncertainty maps for both ground truth and trained predictions, especially in occluded regions.
Evaluation should include semantic IoU, occupancy IoU, uncertainty calibration, occlusion slices, and disagreement between evidence and prediction.
Do not compare only semantic argmax maps; the evidential contribution is the uncertainty and conflict representation.

Treats occupancy labels as uncertain measurements instead of perfect truth.
Handles noisy and partially annotated LiDAR data more honestly than hard voxelization.
The general occupied hypothesis is useful when geometry is visible but semantics are unknown.
Provides uncertainty targets for camera-only occupancy models, not only uncertainty heuristics after training.
Makes occluded and contradictory areas explicit in the output.
Supports safety cases where "unknown" must be distinguishable from "free."

Belief assignments depend on sensor hyperparameters and ray-model assumptions.
LiDAR shadowing, multipath, rain, spray, and reflective surfaces can still create misleading evidence.
Camera models trained on evidential targets can learn dataset-specific uncertainty patterns instead of causal uncertainty.
Evidence-theory outputs require downstream consumers that understand belief and uncertainty fields.
A conservative uncertainty map can reduce availability if thresholds are not calibrated to the ODD.
Sparse semantic labels still limit class-specific evidence, especially for rare airside objects.

Strong fit because airside operation requires distinguishing free, occupied, unknown, and semantically ambiguous space.
Evidential labels are useful for reflective aircraft surfaces, occlusions under wings, jet bridges, service equipment, cones, chocks, hoses, and personnel near clutter.
The general occupied state is valuable for FOD-like or unknown ramp objects whose exact class is less important than avoiding them.
Camera predictions should preserve uncertainty near floodlights, wet pavement, rain, de-icing mist, and aircraft glare.
Use uncertainty to increase buffers and trigger teleoperation rather than only to suppress detections.
Airside validation needs independent 3D ground truth or manual review because LiDAR-derived evidence can be wrong around metal structures.

Keep belief masses or uncertainty channels through the perception-planning interface.
Version LiDAR ray-casting parameters, voxel size, class taxonomy, and evidence hyperparameters with every generated label set.
Add an explicit "unknown occupied" channel if the planner can use it.
Check calibration between camera predictions and evidential uncertainty using held-out adverse-condition logs.
Inspect disagreement cases where the semantic argmax is correct but uncertainty is high, and where uncertainty is low but geometry is wrong.
Use uncertainty-triggered data collection for repeated high-uncertainty airport locations.