PRETRAIN.md

Pretrain DETR

To expedite the training process, we’ll begin by pretraining the DETR component of the model. Typically, training the DETR model on a specific dataset (like DanceTrack, SportsMOT, etc.) doesn’t take too long (approximately a few hours).

DETR Weights on COCO

💾 Similar to many other methods (e.g., MOTR and MeMOTR), we also use COCO pretrained DETR weights for initialization. You can obtain them from the following link:

• Deformable DETR: [official repo] [our drive]
• DAB-Deformable DETR: [official repo]

Pretrain DETR on Specific Datasets

Scripts

All our pre-train scripts follows the template script below. You'll need to fill the <placeholders> according to your requirements:

python -m torch.distributed.run --nproc_per_node=8 main.py --mode train --use-distributed True --use-wandb False --config-path <config file path> --data-root <DATADIR> --outputs-dir <outputs dir>

For example, you can pre-train a Deformable-DETR model on DanceTrack as follows:

python -m torch.distributed.run --nproc_per_node=8 main.py --mode train --use-distributed True --use-wandb False --config-path ./configs/pretrain_r50_deformable_detr_dancetrack.yaml --data-root ./datasets/ --outputs-dir ./outputs/pretrain_r50_deformable_detr_dancetrack/

Gradient Checkpoint

Please referring to here to get more information.

Pretrained Weights

💾 You can directly download the pretrained DETR weights we used in our experiments from Google Drive ☁️. Then put them into ./pretrains/ directory.