QUICK_START.md

[Pending] Replace the file with modified one in the original repositories.

GAN training

You can try to generate synthetic images with GAN. That means you need additional training on the GAN. We have a few pending selections: DC-GAN, VAE-GAN and VAE-WGANGP. There are some references(ref1, ref2) for robust training. Unfortunately we don't have diffusion model at this moment.

In a general scenario, you can simply execute the training script:

python modification_gan/synthetic_main.py --ngf 256 --ndf 64 --ema

ReID(Image Retrival) training

Although some checkpoints are available, you are still advised to train your Re-ID model with Market1501. Due to privacy issue, some datasets such as DukeMTMC are no longer open to the public and not acceptable to the academy as well. In the reid folder, you can see how we build the model as well train the model. There are a few versions of models.

The main focus of the project is to construct a lite backbone for mobile development and real-time tracking. But still, we include a model zoo, with CNN-based re-id models, and vision transformer based models, where you can access all of them in the backbones folder. We train the model on both image-based dataset and video-based dataset(w/. ground truth), and the scripts can be access under the same folder.

For non-continual image training

python reid/image_reid_train.py --bs 64 --backbone seres18 --accelerate --center_lamda 0.0005 --instance 4 --dataset market1501 --temperature 2. --epochs 120 --epsilon 1.0

For continual image training

python reid/image_reid_train.py --bs 64 --backbone seres18 --accelerate --center_lamda 0.0005 --instance 4 --continual --eps 0.6 --dataset market1501 --temperature 2. --epochs 120

For image training with SIE

python reid/image_reid_train.py --bs 64 --backbone seres18 --accelerate --center_lamda 0.0005 --instance 4 --continual --eps 0.6 --dataset market1501 --sie --temperature 2. --epochs 120

For image-level testing

python reid/image_reid_inference.py --backbone seres18 --ckpt checkpoint/reid_model_xxx.onnx --eps 0.6 --dataset xxx

For image testing with SIE

python reid/image_reid_inference.py --backbone seres18 --ckpt checkpoint/reid_model_xxx.onnx --eps 0.6 --dataset xxx --sie

For video-level training

python reid/video_reid_train.py --crop_factor 1.0

Training with accelerate

accelerate config
CUDA_VISIBLE_DEVICES="0" accelerate launch xxx.py --args

To fit with yolov8_tracking, please copy your model and checkpoints to trackers/strongsort/models and trackers/strongsort/checkpoint, and modify reid_model_factory.py accordingly. If you want to train the Re-ID model with video dataset, please refer to the video_train script.

Tracking evaluation

For the final evaluation, please refer to this Wiki for details. We also have our script file for tracking.