Research map / Label-efficient and robust learning / Recognition and long-tail
RLE: A Unified Perspective of Data Augmentation for Cross-Spectral 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 Random Linear Enhancement (RLE), a unified view of cross-spectral data augmentation as local linear transformations (Lambertian model), with moderate (MRLE) and radical (RRLE) variants, evaluated on VI-ReID with a PCB baseline.
Problem
Cross-spectral (visible-infrared) person re-identification suffers from non-linear modality discrepancy; existing image-level augmentations (grayscale, channel selection, random erasing) are heuristic and lack a physical model of cross-spectral transformation.
Approach
Using the Lambertian reflection model, the VIS-to-NIR transformation is linear within a single material region, so the global non-linearity arises from varying linear factors across materials. Augmentations are unified as mimicking such local linear transformations: MRLE mixes RGB channels with Beta-sampled weights summing to one, and RRLE multiplies random rectangular patches by random linear fac
Key results
- On SYSU-MM01 all-search, Baseline+RLE+RE reaches R-1 75.4 and mAP 72.4, versus 64.5 R-1 for the baseline and 73.5 R-1 for CAJ with mixed augmentation.
- On RegDB, the RLE+RE framework reports VIS-to-IR R-1 92.8 and mAP 88.6, exceeding the prior best listed (DEEN 91.1 / 85.1).
- Adding RLE to DEEN improves SYSU-MM01 all-search R-1 from 74.7 to 76.2 and indoor-search R-1 from 80.3 to 83.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add RLE+RE to the basic ResNet-50 PCB framework (ablation) | ResNet-50 PCB · SYSU-MM01 all-search | R-1: 64.5 -> 75.4 | Helped |
| Add grayscale transformation (ablation) | ResNet-50 PCB · SYSU-MM01 all-search | R-1: 64.5 -> 66.7 | Helped |
| Add random channel selection (ablation) | ResNet-50 PCB · SYSU-MM01 all-search | R-1: 64.5 -> 68.3 | Helped |
| Add grayscale transformation and random channel selection (ablation) | ResNet-50 PCB · SYSU-MM01 all-search | R-1: 64.5 -> 68.6 | Helped |
| Add MRLE (ablation) | ResNet-50 PCB · SYSU-MM01 all-search | R-1: 64.5 -> 70.2 | Helped |
| Add random erasing (ablation) | ResNet-50 PCB · SYSU-MM01 all-search | R-1: 64.5 -> 71.0 | Helped |
| Add RRLE (ablation) | ResNet-50 PCB · SYSU-MM01 all-search | mAP: 62.9 -> 69.1 | Mixed |
| Add both random erasing and RRLE (ablation) | ResNet-50 PCB · SYSU-MM01 all-search | R-1: 64.5 -> 74.2 | Helped |
| Replace CAJ augmentation with RLE+RE (ablation) | ResNet-50 PCB · SYSU-MM01 indoor-search | R-1: 80.7 -> 84.7 | Helped |
| Use RLE+RE instead of the DEEN augmentation setup, removing random grayscale and adding RLE (ablation) | DEEN · SYSU-MM01 all-search | R-1: 74.7 -> 76.2 (+1.5) | Helped |
| Add RLE (ablation) | ViT-B with random erasing · SYSU-MM01 all-search | R-1: 66.0 -> 70.2 (+4.2) | Helped |
| Set MRLE beta_m to 0.3 rather than 1.0 (ablation) | MRLE · SYSU-MM01 all-search | R-1: 67.9 -> 70.2 | Helped |
Surprising findings
- Radical RRLE alone scores lower mAP than RE alone on SYSU-MM01 all-search (69.1 vs 69.5), though its R-1 is higher.
Limitations
- The Lambertian model may not hold under extreme weather such as heavy rain, fog, or low illumination.
- Evaluated only on two relatively small VI-ReID datasets (SYSU-MM01, RegDB), so open-world generalization is unverified.
- RLE may not generalize as broadly as generic augmentations like random flipping to other vision tasks.
Experimental setup
Datasets: SYSU-MM01 RegDB
Benchmarks: SYSU-MM01 RegDB
Metrics: Rank-1 Rank-5 Rank-10 Rank-20 mAP mINP
Baselines: Gray RC RE CAJ DEEN DDAG MMN DART ViT-B
Builds on: Lambertian reflection model random erasing PCB DEEN
Closest papers by meaning
- Multi-Modal Object Re-identification via Sparse Mixture-of-Experts (ICML 2025 · similarity 0.83)
- Cross-Modality Perturbation Synergy Attack for Person Re-identification (NeurIPS 2024 · similarity 0.81)
- MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification (NeurIPS 2025 · similarity 0.81)
- Beyond Single-Modal Boundary: Cross-Modal Anomaly Detection through Visual Prototype and Harmonization (CVPR 2025 · similarity 0.81)
- Diffusion-based Synthetic Data Generation for Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.81)
- UCM-VeID V2: A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification (CVPR 2025 · similarity 0.80)
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.