DeepAccident-360: Accident-Centric BEV Benchmark

• DeepAccident-360 is a benchmark dataset offering fully annotated BEV maps from a single 360° panoramic image for accident-centric driving scenarios.
• It utilizes a sophisticated simulation pipeline with six camera inputs and panoramic stitching to generate detailed top-down semantic grids over a 100m x 100m region.
• The dataset challenges segmentation algorithms with realistic occlusions, debris, and class imbalances, evaluated using mIoU and focal loss to highlight system robustness.

DeepAccident-360 is a large-scale benchmark dataset designed for Bird's-Eye-View (BEV) semantic mapping from a single 360° panoramic image, specifically targeting accident-centric driving scenarios in simulated but photorealistic environments. Developed as part of the OneBEV project, DeepAccident-360 extends the DeepAccident CARLA simulation pipeline to address the challenges inherent to dense scene understanding during vehicular accidents, encompassing both static and dynamic semantics over 17 rigorously defined classes. Providing fully annotated top-down (BEV) semantic maps derived directly from simulator metadata, DeepAccident-360 serves as a challenging evaluation and training resource for panoramic BEV segmentation algorithms in abnormal, occlusion-rich, and physically dynamic environments (Wei et al., 2024).

1. Motivation and Design Objectives

DeepAccident-360 was created to furnish the autonomous driving community with an accident-centric BEV mapping benchmark, emphasizing rare and hazardous driving scenarios. The dataset targets comprehensive coverage of scene elements during accident events (such as collisions, multi-vehicle blockages, and obstructions), where road geometry, object presence, and semantic relationships may be substantially altered or partially occluded due to debris or atypical object placement. It is tailored for the one-camera, 360° panoramic BEV segmentation task, intentionally diverging from typical multi-sensor or narrow-field settings by exposing models to highly variable, non-standard visual distributions.

Key objectives include:

• Dense, high-resolution semantic annotation over a full $100\,\mathrm{m} \times 100\,\mathrm{m}$ BEV region around the ego-vehicle.
• Balanced representation of both static (road, buildings, bridges, guard rails) and dynamic (vehicles, pedestrians) scene elements.
• Realistic accident generation by leveraging CARLA simulations, supporting repeatable and well-labeled experimental protocols.

2. Data Generation Pipeline and Sensor Simulation

All scenes are synthesized using CARLA simulator (v0.9.x) in combination with the DeepAccident scenario generation engine, providing varied accident morphologies (e.g., collisions, skidding, occluded or blocked roads). The visual input to each scene is formed by panoramic stitching of images sampled from six pinhole cameras arranged around the vehicle roof, identical in layout to nuScenes. Each imaging device covers approximately $70^\circ$ field-of-view (with the rear camera at $110^\circ$), enabling dense angular coverage.

Panoramic image construction proceeds as follows:

• Each narrow-FOV image is projected onto a tangent sphere of radius $R_i$, where $R_i = r_i + \Delta_i$, and

$$f_x, f_y\, [\mathrm{px}] = \frac{f\,[\mathrm{mm}]}{\mathrm{pixel\_size}} ,\qquad \Delta_i\,[\mathrm{px}] = \sqrt{((C_{i_x} - O_x)\frac{f_x}{f})^2 + ((C_{i_y} - O_y)\frac{f_y}{f})^2}$$

• For each panoramic pixel $P_{360}^{(j,k)}$, the algorithm computes the intersection of the projection ray with the camera's tangent sphere, then warps coordinates back onto the appropriate camera plane (as detailed in the original equations (4)-(7) (Wei et al., 2024)).
• The resultant panoramic images possess a resolution of $600 \times 9600$ pixels (H × W).

3. Annotation Scheme and Label Taxonomy

Semantic ground truth is directly inherited from CARLA’s BEV semantic maps, eliminating manual labeling and ensuring grid-level accuracy. Each $200 \times 200$ BEV cell within $$\left[-50, 50\right]\,\mathrm{m} \times \left[-50, 50\right]\,\mathrm{m}$$ centered on the ego-vehicle receives an integer class label.

From the full DeepAccident labelset, the following 17 semantic classes are retained:

Retained Classes	Static/Dynamic
static, dynamic, building, fence, water, terrain, pedestrian, pole, road_line, road, side_walk, vegetation, vehicles, wall, ground, bridge, guard_rail	mixed (see note)

“Unlabeled,” “other,” and “sky” are removed; classes with zero pixel ratio or low presence ratio (e.g. traffic_light, traffic_sign, rail_track) are omitted as well.

No front-to-BEV label reprojection is necessary due to the direct sampling from CARLA's BEV semantic layer, which ensures accurate alignment and taxonomy congruence between panorama-derived features and ground-truth output maps.

4. Dataset Composition and Statistical Properties

DeepAccident-360 comprises 587 unique accident scenes: 483 for training and 104 for validation. In total, the dataset contains 48,812 panoramic frames (40,619 train / 8,193 val). Each sample consists of a $70^\circ$0 panoramic image paired with a $70^\circ$1 BEV semantic grid.

The dataset does not provide a held-out test set; researchers are expected to rely on the prescribed train/val split for model validation and reporting.

Critical statistical attributes include:

• Number of semantic classes: 17
• BEV grid coverage: $70^\circ$2 domain
• Rare classes (“pedestrian,” “bridge,” “fence,” “guard_rail”) exhibit presence ratios below 20% and occupy a minimal fraction of total BEV pixels
• Occlusions and debris, as well as stitching artifacts, routinely degrade pixel-level clarity for certain classes (notably “road_line,” “side_walk”)

5. Label Distribution, Scene Challenges, and Artifacts

Class-wise pixel and presence ratios show substantial imbalance, with certain infrastructure categories such as “bridge” and “guard_rail,” and most dynamic classes (notably “pedestrian”), appearing infrequently. The simulation of accident scenarios increases occlusion rates, generates irregular vehicle and debris patterns, and disrupts canonical class boundaries. These conditions contribute significant label noise, especially for road-adjacent classes (side_walk, road_line).

The panoramic stitching algorithm introduces further challenges, including seam artifacts, ghosting, and spatial warping, particularly at camera boundaries. These distortions complicate both visual feature extraction and semantic segmentation tasks, particularly for rare or partially visible objects. As a result, DeepAccident-360 constitutes an especially stringent testbed for robust, generalizable BEV mapping from panoramic single-view input.

6. Evaluation Metrics and Loss Functions

Performance on DeepAccident-360 is reported primarily using mean Intersection over Union (mIoU), defined over $70^\circ$3 classes as:

$70^\circ$4

where $70^\circ$5 and $70^\circ$6 denote the predicted and ground-truth BEV masks for class $70^\circ$7.

During training, focal loss is applied on the BEV grid to address pronounced class imbalance and emphasize difficult-to-predict samples.

7. Benchmarking: Model Performance and Observations

DeepAccident-360 has been used to benchmark several leading panoramic BEV segmentation algorithms. Validation mIoU and per-class Intersection over Union are summarized as follows (selected classes shown for brevity):

Method	#Param	static	dynamic	road	vehicles	bridge	pedestrian	mIoU
HDMapNet	12.8M	12.7	15.4	73.4	15.2	1.4	0.0	28.8
BEVFusion	53.7M	10.0	11.1	66.4	13.2	0.0	0.0	23.8
BEVSegformer	52.8M	13.6	17.6	71.6	22.8	0.5	0.0	30.8
BEVFormer	42.9M	10.5	11.6	69.0	19.7	0.0	0.0	26.1
360BEV	32.0M	15.6	16.3	76.7	29.5	5.4	0.0	34.3
OneBEV	31.0M	17.5	19.6	80.1	30.4	4.7	0.0	36.1

Notably, OneBEV achieves the highest overall mIoU (36.1), with pronounced improvements for “water,” “terrain,” and “road,” classes less affected by accident dynamics. Performance for dynamic and rare classes (vehicles, pedestrians, bridge, guard_rail) remains low; for “pedestrian,” all models report 0 IoU, attributed to both low scenario frequency and extreme occlusion or absence after panoramic fusion.

Qualitative evaluation reveals stitched panorama artifacts often disrupt fine boundary delineation and that challenging environmental perturbations from accident generation exacerbate inherent segmentation difficulties. A plausible implication is the potential benefit of targeted synthetic over-sampling or scenario generation for rare/rarely visible infrastructure classes.

Overall, DeepAccident-360 poses a domain-specific, high-fidelity challenge for panoramic BEV semantic mapping, supporting the development and evaluation of algorithms capable of handling complex, dynamically evolving, and artifact-prone driving environments (Wei et al., 2024).