Research map / Detection and segmentation / 3D detection and driving
SparseAlign: a Fully Sparse Framework for Cooperative 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 SparseAlign, a fully sparse framework for LiDAR cooperative object detection built on a sparse UNet with coordinate-expandable convolutions, query-based temporal alignment, pose alignment, and sparse spatial fusion, achieving state-of-the-art accuracy with under 1.3 MB bandwidth.
Problem
Cooperative 3D object detection methods rely on dense BEV feature maps that are computationally expensive, scale poorly to long range, and require high communication bandwidth; fully sparse alternatives suffer from center feature missing and isolated convolution fields.
Approach
Replace dense BEV features with a fully sparse pipeline: a Sparse UNet using coordinate-expandable sparse convolutions (CEC) in 3D and 2D to fix center-feature missing and isolated receptive fields, plus query-based temporal alignment (TAM), neighborhood-graph pose alignment (PAM), sparse query fusion (SAM), and a CompassRose orientation encoding.
Key results
- On OPV2V, SparseAlign reaches AP0.7 0.892 (AP0.5 0.930) with under 1.3 MB bandwidth, approaching V2VFormer++ (L+C) at 0.895.
- On DAIR-V2X, SparseAlign achieves AP0.5 0.845 and AP0.7 0.685, outperforming DI-V2X (0.788 / 0.662).
- On DAIR-V2Xt under TA-COOD, SparseAlign reaches AP0.7 0.548 versus StreamLTS 0.404, a 14.4-point gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SparseAlign instead of DI-V2X | cooperative detection · DAIR-V2X | AP0.5: 0.788 -> 0.845 | Helped |
| use SparseAlign instead of CoAlign | cooperative detection · DAIR-V2X | AP0.7: 0.604 -> 0.685 | Helped |
| use SparseAlign instead of StreamLTS + SA | time-aligned cooperative detection · OPV2Vt | AP0.7: 0.790 -> 0.818 | Helped |
| use SparseAlign instead of StreamLTS + SA | time-aligned cooperative detection · DAIR-V2Xt | AP0.7: 0.497 -> 0.548 | Helped |
| add SAM instead of omitting spatial fusion (ablation) | SparseAlign · OPV2V | AP0.5: 0.847 -> 0.940 | Helped |
| add SAM instead of QUEST spatial fusion (ablation) | SparseAlign · OPV2V | AP0.7: 0.869 -> 0.901 | Helped |
| replace standard sparse convolutional layers with CEC layers (ablation) | SparseAlign SUNet · OPV2V | AP0.7: 0.901 -> 0.929 | Helped |
| add TAM using historical frames (ablation) | SparseAlign · OPV2V | AP0.7: 0.914 -> 0.929 | Helped |
| use CompassRose instead of direct ground-truth angle regression (ablation) | SparseAlign detection head · OPV2V | AP0.7: 0.768 -> 0.885 | Helped |
| use CompassRose instead of sine-cosine encoding (ablation) | SparseAlign detection head · DAIR-V2X | AP0.7: 0.603 -> 0.638 | Helped |
| remove TAM (ablation) | SparseAlign · OPV2Vt at 200ms latency | AP0.7: 0.765 -> 0.209 | Hurt |
| remove point-wise timestamps (ablation) | SparseAlign · OPV2Vt at 200ms latency | AP0.7: 0.765 -> 0.581 | Hurt |
Surprising findings
- Removing training latency (TL) gives higher AP0.7 on OPV2Vt (0.843) than the full model (0.818).
Limitations
- Performance drops about 10% AP under large pose errors even with pose correction.
- Evaluated only on LiDAR-based vehicle detection in OPV2V/DAIR-V2X variants; roadside benchmarks not explored.
Experimental setup
Datasets: OPV2V DAIR-V2X OPV2Vt DAIR-V2Xt
Benchmarks: OPV2V DAIR-V2X OPV2Vt DAIR-V2Xt
Metrics: AP@0.5 AP@0.7 communication bandwidth (MB)
Baselines: Fcooper V2VNet FPVRCNN OPV2V CoAlign CoBEVT V2VAM V2VFormer++ V2XViT DiscoNet DI-V2X StreamLTS
Builds on: MinkUnet StreamLTS CoAlign QUEST Fcooper
Closest papers by meaning
- CorrBEV: Multi-View 3D Object Detection by Correlation Learning with Multi-modal Prototypes (CVPR 2025 · similarity 0.88)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.87)
- DI-V2X: Learning Domain-Invariant Representation for Vehicle-Infrastructure Collaborative 3D Object Detection (AAAI 2024 · similarity 0.87)
- Instance-Aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning (AAAI 2024 · similarity 0.86)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · similarity 0.85)
- BEVSync: Asynchronous Data Alignment for Camera-based Vehicle-Infrastructure Cooperative Perception Under Uncertain Delays (AAAI 2025 · 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.