🌟 The codes for our CVPR 2023 paper 'DropMAE: Masked Autoencoders with Spatial-Attention Dropout for Tracking Tasks'. [Link]

If you find our work useful in your research, please consider citing:

@inproceedings{dropmae2023,
  title={DropMAE: Masked Autoencoders with Spatial-Attention Dropout for Tracking Tasks},
  author={Qiangqiang Wu and Tianyu Yang and Ziquan Liu and Baoyuan Wu and Ying Shan and Antoni B. Chan},
  booktitle={CVPR},
  year={2023}
}

• DropMAE leverages more temporal cues for reconstruction.

• Pre-training Code
• Pre-trained Models
• Fine-tuning Code for VOT
• Fine-tuned Models for VOT
• Fine-tuning Code for VOS
• Fine-tuned Models for VOS

• This repo is a modification based on the MAE repo. Installation follows that repo. You can also check our requirements file.

• In the dropmae pre-training, we mainly use Kinetics Datasets, which can be download in this Link. We use its training raw videos (*.mp4) for training. The detailed download instruction can also be found here.

To pre-train ViT-Base (the default configuration) with multi-node distributed training, run the following on 8 nodes with 8 GPUs each:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=8 \
--node_rank=$INDEX --master_addr=$CHIEF_IP --master_port=1234  main_pretrain_kinetics.py --batch_size 64 \
--model mae_vit_base_patch16 \
--norm_pix_loss \
--mask_ratio 0.75 \
--epochs 400 \
--warmup_epochs 40 \
--blr 1.5e-4 \
--weight_decay 0.05 \
--P 0.1 \
--frame_gap 50 \
--data_path $data_path_to_k400_training_videos \
--output_dir $output_dir \
--log_dir $log_dir

• Here the effective batch size is 64 (batch_size per gpu) * 8 (nodes) * 8 (gpus per node) = 4096. If memory or # gpus is limited, use --accum_iter to maintain the effective batch size, which is similar to MAE.
• P is the spatial-attention dropout ratio for DropMAE.
• data_path indicates the Kinetics (e.g., K400 and K700) training video folder path.
• The exact same hyper-parameters and configs (initialization, augmentation, etc.) are used in our implementation w/ MAE.

The pre-training logs of K400-1600E and K700-800E are provided.

• We also provide the pre-trained models (ViT-Base) on K400 and K800 datasets.
• Conviniently, you could try your tracking model w/ our pre-trained models as the initialization weights for improving downstream performance.

• The OSTrack w/ our DropMAE pre-trained models can achieve state-of-the-art performance on existing popular tracking benchmarks.

• The detailed fine-tuning codes && models can be found in our DropTrack repository.

• The detailed VOS fine-tuning can be found in our DropSeg repository.