Research map / Label-efficient and robust learning / Self-supervised and contrastive

UCM-VeID V2: A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification

CVPR 2025 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Pretext-task visual pre-training · 3 citations (OpenAlex)

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 UCM-VeID V2, a larger UAV RGB-IR vehicle re-ID dataset, and CPSL, a self-supervised cross-modality pre-training method (patch-mixed reconstruction, modality adversarial learning, contrastive learning) that replaces ImageNet initialization to reduce modality bias.

Problem

Cross-modality (RGB-IR) vehicle re-identification from UAVs is limited by small datasets, and ImageNet-pretrained ResNet50 initialization causes modality bias training (MBT) that hurts IR discrimination.

Approach

Pre-train a ResNet50 encoder self-supervisedly on cross-modality ImageNet (RGB images plus CycleGAN-generated IR counterparts) using random RGB-IR patch swapping with patch-mixed reconstruction, modality-discrimination adversarial learning, and contrastive learning between paired mixed features, then fine-tune VT-ReID models from these weights.

Key results

• CPSL pre-training with AGW raises single-shot IR2RGB rank-1 from 15.75% (ImageNet-1K) to 23.47% and mAP from 25.74% to 33.69% on UCM-VeID V2.
• Ablation: adding PMIR, MDAL and MACL together yields rank-1 25.36% and mAP 36.57% in single-shot IR2RGB, versus 16.17% and 26.54% for ImageNet-1K initialization.
• Among SSL baselines, CPSL reaches rank-1 32.64% and mAP 24.01% on multi-shot, outperforming MoCo V2 (28.89% rank-1, 22.44% mAP).

Measured effects reported in the paper

Surprising findings

• Randomly initialized VT-ReID models generally fail to converge, which the authors attribute to triplet-loss gradient explosion.
• ImageNet-pretrained weights perform worse on the larger UCM-VeID V2 than on the original UCM-VeID, a modality-bias effect.

Limitations

• Fine-tuning hyperparameters were not tuned for CPSL, so results may be suboptimal.
• Pre-training relies on CycleGAN-generated IR images, which may not match real infrared statistics.
• Evaluation is limited to a single dataset (UCM-VeID V2).

Experimental setup

Datasets: ImageNet UCM-VeID V2

Benchmarks: UCM-VeID V2

Metrics: rank-1 rank-10 rank-20 mAP CMC

Baselines: AGW DDAG MMD-ReID DEEN MMGL Jigsaw RotationNet InstDisc SimCLR SwAV BYOL SimSiam

Builds on: CycleGAN AGW MoCo ResNet50

Closest papers by meaning

• MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification (NeurIPS 2025 · similarity 0.86)
• Multi-Modal Object Re-identification via Sparse Mixture-of-Experts (ICML 2025 · similarity 0.85)
• Cross-Modality Perturbation Synergy Attack for Person Re-identification (NeurIPS 2024 · similarity 0.84)
• Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.84)
• Implicit Discriminative Knowledge Learning for Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.84)
• Magic Tokens: Select Diverse Tokens for Multi-modal Object Re-Identification (CVPR 2024 · similarity 0.83)