JeSOS ReID

ReID

ReID + Kalman

Person Re-Identification (ReID) is a lightweight deep learning model that specializes in learning feature representations of a person.

Here, the person's features are represented as a heatmap applied to the body. An activation map is computed by taking the sum of absolute-valued feature maps along the channel dimension, followed by a spatial L2 normalization.

How the engine works

A video of a person is fed to the ReID model to get the features, which are used for comparison with others. The L2 Distance is used for feature matching.

$$ L2\left( p,q\right) = \frac{1}{n} \sqrt {\sum _{i=1}^{n} \left( p_i-q_i \right)^2 } $$

• p = customer's features
• q = features to be compared with customer's features

Install requirements

./setup_env.sh

ReID model training on custom dataset

1. Make a folder where video data will be uploaded with the following folder architecture

.path/to/video/folder/

├── person_1_video
├── person_2_video
├── ...
├── ...
└── person_n_video

2. Train model with train.py

Training usage

python3 train.py --name osnet_x1_0 --videos_path /path/to/model --aug_count 5 --save_path /path/to/save

• Model name: --name osnet_x1_0
• Image height: --img_h 256
• Image width: --img_w 128
• Batch size: --bs 32
• Optimizer: --optim adam
• Learning rate: --lr 0.003
• Learning rate scheduler: --lr_sch single_step
• Step size for learning rate scheduler: --step 10
• Epochs: --epochs 30
• Evaluation frequency: --eval_freq 10
• Video data folder path: --videos_paths path/to/folder
• Take every N-th frame from the video: --skip_frames 15
• Augmentations count: --aug_count 7
• Path to save data: --save_path path/to/save

Show help message for all options:

python3 train.py -h

Performance of OSNet models

Model name	Jetson Nano	RTX 2080ti	RTX 3090ti
osnet_x0_25	50ms	12.1ms	9.1ms
osnet_x0_5	53.4ms	12.5ms	9.5ms
osnet_x0_75	55.3ms	13ms	10.2ms
osnet_x1_0	59.8ms	13.5ms	11ms

Convert model to ONNX

1. Install ONNX

pip3 install onnx

2. Convert using reid_to_onnx.py script

python3 reid_to_onnx.py --name osnet_x1_0 --nc 3

Usage

• Model name: --name osnet_x1_0
• Number of classes: --nc 3
• Path to weights: --weights path/to/model.pt
• Image height: --img_h 256
• Image width: --img_w 128 Show help message for all options:

python3 reid_to_onnx.py -h

Convert the model to TensorRT

1. To convert the ReID model to TensorRT, you should first convert the model to ONNX.

2. Then, install TensorRT Backend For ONNX onnx-tensorrt library.

3. Run the script below.

onnx2trt osnet_x1_0.onnx -o osnet_x1_0.trt

Docker

Make sure to have nvidia-docker installed.

docker run --runtime=nvidia -it --volume /home/ec2-user/reid:/opt/project reid-dev:gpu /bin/bash

Common issues

ImportError: libcudnn.so.7: cannot open shared object file: No such file or directory

Run this command in the terminal:

echo "export LD_LIBRARY_PATH=/usr/local/cuda/lib64:\${LD_LIBRARY_PATH}" >> ${HOME}/.bashrc