Research map / Detection and segmentation / 3D detection and driving
DI-V2X: Learning Domain-Invariant Representation for Vehicle-Infrastructure Collaborative 3D Object Detection
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.
Proposes DI-V2X, a teacher-student distillation framework that learns domain-invariant BEV features for V2X intermediate-fusion 3D detection across vehicle and infrastructure LiDARs, using domain-mixing instance augmentation, progressive distillation, and domain-adaptive fusion.
Problem
V2X collaborative 3D object detection suffers from a data-level domain gap because vehicles and infrastructure use different LiDAR sensors, which degrades fused performance.
Approach
An early-fused teacher on merged point clouds guides student models from each domain via two-stage distillation (non-overlapping regions before fusion, overlapping regions after fusion), combined with instance-level mixing of domain sources and a fusion module that corrects calibration offsets and applies domain- and spatially-adaptive attention.
Key results
- On DAIR-V2X val, DI-V2X student reaches 78.82 BEV AP@0.5 and 66.16 AP@0.7, beating CoAlign by 4.22 and 5.76 points.
- On V2XSet, DI-V2X student reaches 92.7 AP@0.5 and 82.7 AP@0.7, +11.5 AP@0.7 over V2X-ViT.
- Ablation on DAIR-V2X: DMA, PDD, and DAF together raise PP-IF from 72.85/52.67 to 78.82/66.16 (AP@0.5/AP@0.7).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DI-V2X instead of CoAlign | PointPillars intermediate-fusion detection on DAIR-V2X · DAIR-V2X val | BEV AP@0.5: 74.60 -> 78.82 (+4.22) | Helped |
| Use DI-V2X instead of CoAlign | PointPillars intermediate-fusion detection on DAIR-V2X · DAIR-V2X val | BEV AP@0.7: 60.40 -> 66.16 (+5.76) | Helped |
| Use DI-V2X instead of V2X-ViT | PointPillars intermediate-fusion detection on V2XSet · V2XSet val | AP@0.5: 88.2 -> 92.7 (+4.5) | Helped |
| Use DI-V2X instead of V2X-ViT | PointPillars intermediate-fusion detection on V2XSet · V2XSet val | AP@0.7: 71.2 -> 82.7 (+11.5) | Helped |
| Add DMA to PP-IF (ablation) | PointPillars intermediate-fusion detection · DAIR-V2X val | AP@0.5: +1.05% | Helped |
| Add DMA to PP-IF (ablation) | PointPillars intermediate-fusion detection · DAIR-V2X val | AP@0.7: +1.35% | Helped |
| Add PDD to PP-IF (ablation) | PointPillars intermediate-fusion detection · DAIR-V2X val | AP@0.5: 1.32% | Helped |
| Add PDD to PP-IF (ablation) | PointPillars intermediate-fusion detection · DAIR-V2X val | AP@0.7: 3.22% | Helped |
| Add DAF to PP-IF (ablation) | PointPillars intermediate-fusion detection · DAIR-V2X val | AP@0.5: +2.91% | Helped |
| Add DAF to PP-IF (ablation) | PointPillars intermediate-fusion detection · DAIR-V2X val | AP@0.7: +8.22% | Helped |
| Add DMA, PDD, and DAF together (ablation) | PointPillars intermediate-fusion detection · DAIR-V2X val | AP@0.5: 72.85 -> 78.82 | Helped |
| Add DMA, PDD, and DAF together (ablation) | PointPillars intermediate-fusion detection · DAIR-V2X val | AP@0.7: 52.67 -> 66.16 | Helped |
Surprising findings
- Late fusion (PP-LF) lowers AP@0.5 relative to no fusion on DAIR-V2X (63.68 vs 65.21), attributed to pose error.
- The early-fused teacher is the best single model in the domain generalization test on DAIR-V2X, yet the DI-V2X student still outperforms all intermediate fusion baselines.
Limitations
- Evaluated only on two V2X datasets (DAIR-V2X and V2XSet), with the teacher requiring early-fused point clouds during training.
- Gains on DAIR-V2X infrastructure-only testing are below the infrastructure-trained PP baseline (45.91 vs 65.80 for Inf.-only model).
Experimental setup
Datasets: DAIR-V2X V2XSet
Benchmarks: DAIR-V2X V2XSet
Metrics: BEV AP@0.5 BEV AP@0.7 AP
Baselines: PP-LF PP-EF PP-IF PP-KD PP-CDT V2VNet DiscoNet OPV2V Where2comm CoAlign V2X-ViT
Builds on: PointPillars VoxelNet CoAlign DiscoNet mixup
Closest papers by meaning
- Triple Feature Disentanglement for One-Stage Adaptive Object Detection (AAAI 2024 · similarity 0.87)
- DSRC: Learning Density-Insensitive and Semantic-Aware Collaborative Representation Against Corruptions (AAAI 2025 · similarity 0.87)
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.87)
- Learning to Detect Objects from Multi-Agent LiDAR Scans without Manual Labels (CVPR 2025 · similarity 0.87)
- DiffuBox: Refining 3D Object Detection with Point Diffusion (NeurIPS 2024 · similarity 0.87)
- Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.87)
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.