Research map / Label-efficient and robust learning / Recognition and long-tail
LiDAR-based Person Re-identification
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 ReID3D, the first LiDAR-based person re-identification framework, with a multi-task simulated pre-training scheme and a graph-based complementary encoder, plus a new real LiDAR ReID dataset (LReID) and a simulated dataset (LReID-sync).
Problem
Camera-based person re-identification suffers from poor illumination, cluttered backgrounds, appearance variation and privacy concerns, and prior 3D sensors (Kinect, radar) are limited in range or scale; LiDAR-based person ReID had not been studied.
Approach
Use LiDAR point clouds for ReID with a graph-convolution encoder (GCEE) containing a Complementary Feature Extractor with an Eraser module, pre-trained on simulated multi-view pedestrian point clouds for point cloud completion and SMPL shape regression, followed by a transformer temporal module.
Key results
- ReID3D reaches 94.0 Rank-1, 98.0 Rank-3 and 83.28 mAP on LReID overall, versus 87.0 Rank-1 for the best camera baseline TCLNet.
- Under low light, ReID3D achieves 93.3 Rank-1 versus 60.0 for TCLNet.
- Multi-task pre-training (completion + SMPL) raises Rank-1 from 90.0 without pre-training to 94.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| multi-task pre-training (point cloud completion + SMPL shape) on LReID-sync (ablation) | ReID3D · LReID | Rank-1: 90.0 to 94.0 | Helped |
| completion-only pre-training instead of multi-task (ablation) | B-ReID3D · LReID | Rank-1: 90.0 to 93.0 (completion-only); 94.0 with SMPL added | Helped |
| pre-training with similar ReID task (ablation) | B-ReID3D · LReID | Rank-1: 90.0 to 91.0 | Helped |
| Graph-based Complementary Enhancement Encoder (GCEE) instead of PointNet (ablation) | B-ReID3D · LReID | Rank-1: 74.0 (PointNet) to 90.0 | Helped |
| remove Complementary Feature Extractor (CFE) (ablation) | B-ReID3D · LReID | Rank-1: 90.0 to 83.0 | Hurt |
| remove Eraser module from CFE (ablation) | B-ReID3D · LReID | Rank-1: 90.0 to 85.0 | Hurt |
| transformer temporal module instead of LSTM (ablation) | B-ReID3D · LReID | Rank-1: 82.0 (LSTM) to 90.0; 80.0 (Bi-LSTM) | Helped |
| LiDAR point clouds instead of camera video (ReID3D vs camera methods) | ReID3D · LReID low light | Rank-1: 93.3 vs 60.0 (TCLNet best camera method) | Helped |
| LiDAR point clouds instead of camera video (ReID3D vs camera methods) | ReID3D · LReID normal light | Rank-1: 94.3 vs 98.6 (TCLNet) | Hurt |
| LiDAR point clouds instead of camera video (ReID3D vs camera methods) | ReID3D · LReID overall | Rank-1: 94.0 vs 87.0 (TCLNet) | Helped |
Surprising findings
- Camera-based video methods outperform ReID3D under normal light (e.g., TCLNet 98.6 vs 94.3 Rank-1), attributed to appearance cues being unavailable to LiDAR.
Limitations
- Falls behind camera-based video methods under normal lighting where appearance cues are available.
- Evaluation limited to a single LiDAR-based dataset collected in two outdoor scenes with 100 test identities.
Experimental setup
Datasets: LReID LReID-sync
Benchmarks: LReID
Metrics: Rank-1 Rank-3 mAP CMC
Baselines: TCLNet STMN SINet PiT PointNet Point Transformer 3DCNN LSTM Bi-LSTM
Builds on: PCN TCLNet Point Transformer batch-hard triplet loss
Closest papers by meaning
- Clothing-Change Feature Augmentation for Person Re-Identification (CVPR 2023 · similarity 0.80)
- Diverse Embedding Expansion Network and Low-Light Cross-Modality Benchmark for Visible-Infrared Person Re-Identification (CVPR 2023 · similarity 0.80)
- Person De-reidentification: A Variation-guided Identity Shift Modeling (CVPR 2025 · similarity 0.78)
- Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification (NeurIPS 2024 · similarity 0.78)
- FlexiReID: Adaptive Mixture of Expert for Multi-Modal Person Re-Identification (ICML 2025 · similarity 0.78)
- SeCap: Self-Calibrating and Adaptive Prompts for Cross-view Person Re-Identification in Aerial-Ground Networks (CVPR 2025 · similarity 0.78)
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.