Research map / Detection and segmentation / 3D detection and driving
DSRC: Learning Density-Insensitive and Semantic-Aware Collaborative Representation Against Corruptions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces a benchmark of six natural corruptions for multi-agent LiDAR collaborative perception and proposes DSRC, a teacher-student sparse-to-dense, semantic-painted distillation framework with point cloud reconstruction for robust 3D detection.
Problem
Collaborative LiDAR 3D object detection degrades under natural corruptions (weather, sensor failures, external disturbances), and no robustness benchmark existed for multi-agent collaborative perception.
Approach
A teacher trained on multi-view dense point clouds with ground-truth-box semantic painting guides a sparse single-view student via three-stage distillation (encoder, fusion, prediction), with an auxiliary feature-to-point-cloud reconstruction loss to improve cross-agent fusion; only the student is used at inference.
Key results
- On OPV2V, DSRC achieves 92.58/88.45 AP@0.5/0.7 on clean data, outperforming CoBEVT (91.39/86.18) and other intermediate fusion baselines.
- Under corruptions on OPV2V, DSRC reaches 83.54/69.84 under fog versus CoBEVT's 75.04/68.63 at AP@0.5/0.7.
- On DAIR-V2X clean, DSRC reaches 69.51/56.54 AP@0.5/0.7, above CoAlign's 68.23/55.08.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DSRC instead of V2VNet | OPV2V collaborative 3D detection · OPV2V clean | AP@0.5/0.7: 92.29/83.20 -> 92.58/88.45 | Helped |
| Use DSRC instead of CoAlign | DAIR-V2X collaborative 3D detection · DAIR-V2X clean | AP@0.5/0.7: 68.23/55.08 -> 69.51/56.54 | Helped |
| Add sparse-to-dense distillation (ablation) | Detection-only student · OPV2V average across corruptions | AP@0.5/0.7: 79.75/70.54 -> 83.42/73.92 | Helped |
| Add sparse-to-dense distillation (ablation) | Detection-only student · DAIR-V2X average across corruptions | AP@0.5/0.7: 53.11/40.12 -> 53.84/40.98 | Helped |
| Add point cloud painting (ablation) | Detection-only student · OPV2V average across corruptions | AP@0.5/0.7: 79.75/70.54 -> 84.04/74.33 | Helped |
| Add point cloud painting (ablation) | Detection-only student · DAIR-V2X average across corruptions | AP@0.5/0.7: 53.11/40.12 -> 53.97/41.02 | Helped |
| Add point cloud reconstruction to sparse-to-dense distillation and point cloud painting (ablation) | DSRC · OPV2V average across corruptions | AP@0.5/0.7: 84.83/75.99 -> 85.54/76.67 | Helped |
| Add point cloud reconstruction to sparse-to-dense distillation and point cloud painting (ablation) | DSRC · DAIR-V2X average across corruptions | AP@0.5/0.7: 54.28/41.37 -> 54.84/41.98 | Helped |
| Apply DSRC to position-wise maximum fusion (ablation) | Maximum (vanilla) collaborative model · OPV2V clean | AP@0.5/0.7: 87.40/79.40 -> 91.51/86.07 | Helped |
| Apply DSRC to no-collaboration model (ablation) | No Collaboration (vanilla) · DAIR-V2X fog | AP@0.5/0.7: 26.19/19.75 -> 36.32/30.66 | Helped |
Surprising findings
- Most models show minor degradation under crosstalk, attributed to multi-agent settings making such noise present in training data.
- Snow and fog cause average AP@0.7 drops of over 20% across all models.
Limitations
- Motion blur remains a substantial challenge for all models, with performance still degraded.
- Teacher relies on ground-truth bounding boxes for painting during training, so it requires labeled clean data.
Experimental setup
Datasets: OPV2V DAIR-V2X
Benchmarks: OPV2V DAIR-V2X
Metrics: AP@0.5 AP@0.7 mCE mAP
Baselines: No Collaboration Late Fusion F-Cooper V2VNet V2X-ViT CoAlign ERMVP Mrcnet CoBEVT
Builds on: PointPillars OpenCOOD Res2Net Knowledge distillation
Closest papers by meaning
- RCP-Bench: Benchmarking Robustness for Collaborative Perception Under Diverse Corruptions (CVPR 2025 · similarity 0.88)
- DI-V2X: Learning Domain-Invariant Representation for Vehicle-Infrastructure Collaborative 3D Object Detection (AAAI 2024 · similarity 0.87)
- Privacy-Preserving V2X Collaborative Perception Integrating Unknown Collaborators (AAAI 2025 · similarity 0.86)
- Generative Map Priors for Collaborative BEV Semantic Segmentation (CVPR 2025 · similarity 0.86)
- Multi-agent Collaborative Perception via Motion-aware Robust Communication Network (CVPR 2024 · similarity 0.86)
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.