Renhao Wang*, Yu Sun*, Yossi Gandelsman, Xinlei Chen, Alexei A. Efros, Xiaolong Wang

[arXiv] [Project] [BibTeX]

See installation instructions.

We release COCO-Videos, a new dataset for instance and panoptic segmentation which follows the COCO labeling format. We also rely on semantic-level labels in the KITTI-STEP dataset for evaluation on semantic segmentation.
All datasets can be downloaded here and should subsequently be unzipped to the path specified under the $DETECTRON2_DATASETS environment variable (see installation instructions).

Relevant pretrained checkpoints can be obtained here. These should be downloaded and stored at some /path/to/checkpoints.

To evaluate a pretrained Mask2Former-S on COCO-Videos for panoptic segmentation:

python runner_coco_videos_baseline.py --gpu 0 \
  --videos bangkok bar berkeley havana house irvine paris restaurant school tokyo \
  --batch_size 8 \
  --weights /path/to/checkpoints/ttt_coco_panoptic_baseline.pkl \
  --output_dir coco_vid_panoptic_baseline \
  --eval_type pano \
  --num_imgs 4000

You can pass --eval_type inst to obtain the baseline instance numbers (as well as the corresponding pretrained instance segmentation checkpoint). Results will be logged under the directory specified in the --output_dir flag,

Runner script for instance segmentation:

python runner_ttt_mae_inst.py --gpu 0 \
    --videos bangkok bar berkeley havana house irvine paris restaurant school tokyo \
    --batch_size 32 \
    --accum_iter 8 \
    --base_lr 0.0001 \
    --weights /path/to/checkpoints/ttt_coco_instance_baseline.pkl \
    --restart_optimizer

Runner script for panoptic segmentation:

python runner_ttt_mae_panoptic.py --gpu 0 \
    --videos bangkok bar berkeley havana house irvine paris restaurant school tokyo \
    --batch_size 32 \
    --accum_iter 8 \
    --base_lr 0.0001 \
    --weights /path/to/checkpoints/ttt_coco_panoptic_baseline.pkl \
    --restart_optimizer

For easy collation of numbers, we provide a utility script which can, for example, be called as python mask2former/utils/tabulate_results_cv.py --root_dir exp_dir/mae_coco_inst_32_0.0001.

Runner script:

python runner_ttt_mae.py --gpu 0 \
    --videos 0000 0001 0002 0003 0004 0005 0006 0007 0008 0009 0010 0011 0012 0013 0014 0015 0016 0017 0018 0019 0020 \
    --batch_size 32 \
    --accum_iter 4 \
    --base_lrs 0.0001 \
    --weights /path/to/checkpoints/ttt_ks_semantic_baseline.pkl \
    --restart_optimizer

For easy collation of numbers, we provide a utility script which can, for example, be called as python mask2former/utils/tabulate_results.py --root_dir exp_dir/mae_ks_sema_32_0.0001.

This codebase inherits all licenses from the public release of Mask2Former.

@article{wang2023test,
  title={Test-time training on video streams},
  author={Wang, Renhao and Sun, Yu and Gandelsman, Yossi and Chen, Xinlei and Efros, Alexei A and Wang, Xiaolong},
  journal={arXiv preprint arXiv:2307.05014},
  year={2023}
}

Code is based on Mask2Former.