TextReIDNet

This is the codebase for TextReIDNet (Text-based Person Re-identification Network). The details of TextReIDNet are found in our DCOSS-IoT 2024 paper titled "Resource-Efficient Text-based Person Re-identification on Embedded Devices".

 

Summary

TextReIDNet is a lightweight person re-identification model designed explicitly for embedded devices. TextReIDNet has a total of 32.29 million parameters only, but a achieves an impressive 52.76% and 35.71% top-1 accuracy on the CUHK-PEDES and RSTPReid datasets respectively.

Figure 1. The architure of TextReIDNet.

 

Requirements (dependencies)

• Operating System: Ubuntu 20.04.6 LTS (irrelevant but worth mentioning)
• CUDA Version: 11.7
• python version: 3.9.18
• pytorch version: 1.13.1
• torchvision version: 0.14.1
• pillow version: 9.5.0
• opencv-python version: 4.8.1.78
• tqdm version: 4.66.1
• numpy version: 1.26.2
• natsort version: 8.4.0

The remaining requirements are specified in requirements.txt

 

Resources

• CUHK-PEDES dataset: please see here
• RSTPReid dataset: please see here
• Pretrained model (30 epochs): TextReIDNet_State_Dicts.pth.tar

 

Setup

• Install dependencies. Complete list of dependencies is in requirements.txt
• Download or clone TextReIDNet repository
• Navigate into TextReIDNet directory: cd /path/to/TextReIDNet
• Download the pre-trained model TextReIDNet_State_Dicts.pth.tar and put it into TextReIDNet/data/checkpoints/TextReIDNet_State_Dicts.pth.tar
• Read and modify the config.py to suit your system resources and requirements

 

Test (using pretrained model - 30 epochs)

• Please make sure the pre-trained model TextReIDNet_State_Dicts.pth.tar is downloaded put it into TextReIDNet/data/checkpoints/TextReIDNet_State_Dicts.pth.tar
• Also read and modify the config.py to suit your system parameters
• python test.py. The test results are logged to the console only.

 

Inference on sample images

10 sample images are provided in TextReIDNet/data/samples/ for inference purposes. boy_with_bag.jpg and girl_student.jpg are taken from the internet, while the remaining 8 are sampled from the CUHK-PEDES dataset. You may download and crop more single-person images from the internet and add them to TextReIDNet/data/samples/.

• Please make sure the pre-trained model TextReIDNet_State_Dicts.pth.tar is downloaded put it into TextReIDNet/data/checkpoints/TextReIDNet_State_Dicts.pth.tar
• Also read and modify the config.py to suit your system parameters
• modify the value of textual_description in TextReIDNet/inference/search_person.py with the description of the person you would like to retrieve from TextReIDNet/data/samples/.
• run TextReIDNet/inference/search_person.py. The ranking results are logged to the console, while the top-1 image is saved to TextReIDNet/inference/retrieved_image.jpg.

Figure 2. Example of Top-5 person search results on the CUHK-PEDES dataset.

 

Train (from scratch only, no resume capability)

• First read and modify the config.py to suit your system parameters
• Run python train.py. The training progress and values are logged into TextReIDNet/data/logs/train.log

 

Test and Inference on Nvidia Jetson Nano

Download and use the Ubuntu 20.04 OS image from Qengineering to setup the operating system on the Jetson Nano device. The libraries provided by the Ubuntu 20.04 OS image should be sufficient, but where neccessary install the dependencies. The OS image comes pre-installed with the following:

• OpenCV 4.8.0
• Pytorch 1.13.0
• TorchVision 0.14.0
• TensorRT 8.0.1.6

Steps

• Follow the steps outlined here to install torch2trt on the Nano device
• Clone TextReIDNet onto your Nano device
• Download the pre-trained model TextReIDNet_State_Dicts.pth.tar put it into TextReIDNet/data/checkpoints/TextReIDNet_State_Dicts.pth.tar
• Read and modify the TextReIDNet/config.py file (where neccessary)
• Execute TextReIDNet_github/nano/nano_export.py to convert the pytorch model to TensorRT on the Nano
• Execute TextReIDNet_github/nano/nano_import.py to get the run-time values on the Nano

Please note: The initial run-time of the first iteration is always significantly longer than that of subsequent iterations. This is due to the (delayed) loading of additional modules.