Research map / Label-efficient and robust learning / Recognition and long-tail
Catalyst for Clustering-Based Unsupervised Object Re-identification: Feature Calibration
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.
Inserts a parameter-free, training-free nonparametric graph attention feature calibration module before DBSCAN pseudo-label generation in clustering-based unsupervised person ReID, plugged into CC-ViT.
Problem
Clustering-based unsupervised object re-identification suffers from noisy pseudo-labels produced by clustering poor-quality features, which misleads representation learning in a vicious cycle.
Approach
Calibrate encoder features before clustering with a nonparametric graph attention network (cosine-similarity softmax attention, no linear transforms), so similar instances move together and dissimilar ones separate, yielding more reliable pseudo-labels; the module is removed at test time.
Key results
- On MSMT17, CC-FCM-ViT reaches 49.1% mAP versus 40.9% for the CC-ViT baseline (+8.2%), the best reported among compared methods.
- On DukeMTMC-reID, CC-FCM-ViT achieves 71.9% mAP, improving the baseline ViT-small by 2.3%.
- On Market-1501, CC-FCM-ViT reaches 88.6% mAP vs 88.0% for CC-ViT, a modest gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FCM to CC-ViT | CC-ViT · Market-1501 | mAP: 88.0 -> 88.6 | Helped |
| Add FCM to CC-ViT | CC-ViT · MSMT17 | mAP: 40.9 -> 49.1 | Helped |
| Add FCM to CC-ViT | CC-ViT · DukeMTMC-reID | mAP: ViT-small 69.6 -> 71.9 | Helped |
| Set k to 4 instead of 16 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 47.2 | Hurt |
| Set k to 8 instead of 16 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 47.6 | Hurt |
| Set k to 32 instead of 16 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 49.2 | Helped |
| Set k to 64 instead of 16 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 48.0 | Hurt |
| Set temperature t to 0.05 instead of 1 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 47.7 | Hurt |
| Set temperature t to 0.1 instead of 1 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 48.0 | Hurt |
| Set temperature t to 0.5 instead of 1 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 48.3 | Hurt |
| Use 2 SA layers instead of 1 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 48.2 | Hurt |
| Use 3 SA layers instead of 1 (ablation) | CC-FCM-ViT · MSMT17 | mAP: 49.1 -> 48.3 | Hurt |
Surprising findings
- Only one nonparametric SA layer is needed; deeper stacks (2, 3, 5) do not improve mAP consistently.
- Removing learnable transformations from GAT still gives clustering gains, suggesting feature aggregation accounts for most of the benefit.
Limitations
- Gains on Market-1501 are marginal (0.6% mAP).
- Evaluated only on one simple baseline (CC-ViT); not tested on other clustering methods in experiments.
- Extension to unsupervised visible-infrared ReID and domain adaptation is left as future work.
Experimental setup
Datasets: Market-1501 MSMT17 DukeMTMC-reID LUPerson
Benchmarks: Market-1501 MSMT17 DukeMTMC-reID
Metrics: mAP Rank-1 Rank-5 Rank-10 ARI AMI VMS FMS
Baselines: CC-ViT ICE ISE PPLR HNGN U-SSL FPM SpCL IICS JVTC+
Builds on: Graph Attention Networks Cluster Contrast DBSCAN SGC-style decoupled aggregation
Closest papers by meaning
- Relieving Universal Label Noise for Unsupervised Visible-Infrared Person Re-Identification by Inferring from Neighbors (AAAI 2025 · similarity 0.86)
- Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic (ICLR 2024 · similarity 0.86)
- TOP-ReID: Multi-Spectral Object Re-identification with Token Permutation (AAAI 2024 · similarity 0.86)
- Unsupervised Group Re-identification via Adaptive Clustering-Driven Progressive Learning (AAAI 2024 · similarity 0.85)
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.85)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · 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.