TOP-ReID: Multi-spectral Object Re-Identification with Token Permutation

Yuhao Wang · Xuehu Liu · Pingping Zhang* · Hu Lu · Zhengzheng Tu · Huchuan Lu

AAAI 2024 Paper

TOP-ReID is a powerful multi-spectral object Re-identification (ReID) framework designed to retrieve specific objects by leveraging complementary information from different image spectra. It overcomes the limitations of traditional single-spectral ReID in complex visual environments by reducing distribution gap and enhancing cyclic feature aggregation among different image spectra. Besides, TOP-ReID achieves advanced performance in multi-spectral and missing-spectral object ReID and holds great potential under cross-spectral settings.

News

Exciting news! Our paper has been accepted by the AAAI 2024! 🎉 Paper

Table of Contents

• Introduction
• Contributions
• Results
• Visualizations
• Reproduction
• Citation

Introduction

Multi-spectral object ReID is crucial in scenarios where objects are captured through different image spectra, such as RGB, near-infrared, and thermal imaging. TOP-ReID tackles the challenges posed by the distribution gap among these spectra and enhances feature representations by utilizing all tokens of Transformers.

Contributions

• We propose a novel feature learning framework named TOP-ReID for multi-spectral object ReID. To our best knowledge, our proposed TOP-ReID is the first work to utilize all the tokens of vision Transformers to improve the multi-spectral object ReID.
• We propose a Token Permutation Module (TPM) and a Complementary Reconstruction Module (CRM) to facilitate multi-spectral feature alignment and handle spectral-missing problems effectively.
• We perform comprehensive experiments on three multispectral object ReID benchmarks, i.e., RGBNT201, RGBNT100 and MSVR310. The results fully verify the effectiveness of our proposed methods.

Results

Multi-spectral Object ReID

Multi-spectral Person ReID [RGBNT201]

Multi-spectral Vehicle ReID [RGBNT100、MSVR310]

Missing-spectral Object ReID

Missing-spectral Person ReID [RGBNT201]

Missing-spectral Vehicle ReID [RGBNT100]

Performance comparison with different modules [RGBNT201、RGBNT100]

Performance comparison of different backbones [RGBNT201]

Visualizations

T-SNE [RGBNT201]

Grad-CAM [RGBNT201、RGBNT100]

Please check the paper for detailed information Paper

Reproduction

Datasets

RGBNT201 link: https://drive.google.com/drive/folders/1EscBadX-wMAT56_It5lXY-S3-b5nK1wH
RGBNT100 link: https://pan.baidu.com/s/1xqqh7N4Lctm3RcUdskG0Ug code：rjin
MSVR310 link: https://pan.baidu.com/s/1soWMH10DyGG0krTKL7fsYg code: p6ch

Pretrained

ViT-B link: https://pan.baidu.com/s/1YE-24vSo5pv_wHOF-y4sfA
DeiT-S link: https://pan.baidu.com/s/1YE-24vSo5pv_wHOF-y4sfA
T2T-ViT-24 link: https://pan.baidu.com/s/1YE-24vSo5pv_wHOF-y4sfA code: vmfm

Configs

RGBNT201 file: TOP-ReID/configs/RGBNT201/TOP-ReID.yml
RGBNT100 file: TOP-ReID/configs/RGBNT100/TOP-ReID.yml
MSVR310 file: TOP-ReID/configs/MSVR310/TOP-ReID.yml

Bash

#!/bin/bash
source activate (your env)
cd ../(your path)
pip install -r requirements.txt
python train_net.py --config_file ../RGBNT201/TOP-ReID.yml

Training Example

In order to facilitate users in reproducing the results, we have provided training example. It is important to note that there may be slight variations in the experimental results compared to the data presented in the paper. It is worth noting that our model shows significant improvements on the RGBNT201 dataset. This is partly related to the dataset itself and partly to our choice of learning rate. During the experimental process, to align with the learning rate settings in TransReID, we initially adjust the learning rate to 0.008. However, we find that this task is sensitive to the learning rate. When the learning rate is too low, the model's performance fluctuates significantly. Therefore, for better performance and to enhance the competitiveness of our model, we choose a uniform and more suitable learning rate, ultimately selecting 0.009 as the standardized experimental setting. On the smaller MSVR310 dataset, we follow the authors' recommendations, using a higher number of epochs to improve the model's performance. Please note the above details. Here is the example of training TOP-ReID on RGBNT201 and RGBNT100.

RGBNT201:

train.txt

RGBNT100:

train.txt

Tips

If your machine's GPU memory is insufficient, consider adjusting the batch size accordingly. However, be aware that this may potentially impact the results. Moreover, based on our experimental findings, utilizing a single Transformer as the backbone network produces comparable results (mAP: 71.7%, Rank-1: 76.7%) on RGBNT201. To reduce GPU memory usage, you can opt to utilize only one Transformer backbone to process data from all three modalities. This modification requires adjusting the model initialization definition.

Star History

Citation

If you find TOP-ReID useful in your research, please consider citing:

@inproceedings{wang2024top,
  title={TOP-ReID: Multi-spectral Object Re-Identification with Token Permutation},
  author={Wang, Yuhao and Liu, Xuehu and Zhang, Pingping and Lu, Hu and Tu, Zhengzheng and Lu, Huchuan},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={6},
  pages={5758--5766},
  year={2024}
}