As described in the paper, we report the results using the model trained on A2D-Sentences without finetune.
| Backbone | Overall IoU | Mean IoU | mAP | Model |
|---|---|---|---|---|
| Video-Swin-T* | 70.0 | 69.3 | 39.1 | model |
| Video-Swin-T | 71.9 | 71.0 | 42.2 | model |
| Video-Swin-S | 72.8 | 71.5 | 42.4 | model |
| Video-Swin-B | 73.0 | 71.8 | 43.7 | model |
* the model is trained from scratch and set --num_frames 6.
python3 -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --dataset_file jhmdb --with_box_refine --freeze_text_encoder --batch_size 2 --resume [/path/to/model_weight] --backbone [backbone] --eval
For example, evaluating the Video-Swin-Tiny model, run the following command:
python3 -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --dataset_file jhmdb --with_box_refine --freeze_text_encoder --batch_size 2 --resume a2d_video_swin_tiny.pth --backbone video_swin_t_p4w7 --eval