This is the official repository for the following paper:

Chihiro Nakatani, Hiroaki Kawashima, Norimichi Ukita
Interaction-aware Joint Attention Estimation Using People Attributes, ICCV2023
Project page: https://toyota-ti.ac.jp/Lab/Denshi/iim/ukita/selection/ICCV2023-PJAE.html

python 3.6.9

And you can use requirements.txt

pip install -r requirements.txt

You can download daatset from the following url.
These dataset are required to place in data/ in the repository.

• Volleyball dataset (data/videos)
  https://github.com/mostafa-saad/deep-activity-rec

• Volleyball dataset (data/jae_dataset_bbox_gt, data/jae_dataset_bbox_pred)
  https://drive.google.com/drive/folders/1O55_wri92uv87g-2aDh8ll6dFVupmFaB?usp=share_link

• Volleyball dataset (data/vatic_ball_annotation/annotation_data)
  https://drive.google.com/drive/folders/1O55_wri92uv87g-2aDh8ll6dFVupmFaB?usp=share_link

• VideoCoAtt dataset (data/VideoCoAtt_Dataset)
  http:https://www.stat.ucla.edu/~lifengfan/shared_attention

• VideoCoAtt dataset (data/VideoCoAtt_Dataset/dets_heads)
  https://drive.google.com/drive/folders/1O55_wri92uv87g-2aDh8ll6dFVupmFaB?usp=share_link

• You can change parameters of the model (e.g., multi-head numbers, transformer encoder numbers, ...) by editing the yaml files.
• Trained model are also published in here (https://drive.google.com/drive/folders/1O55_wri92uv87g-2aDh8ll6dFVupmFaB?usp=share_link )
• trained models required to place in saved_weights/volleyball or saved_weights/videocoatt in the repository.

• Ours

python train.py yaml/volleyball/train_ours_p_p.yaml
python train.py yaml/volleyball/train_ours.yaml

The following folder contains the trained models.

1. volleyball-dual-mid_p_p_field_middle_p_s_davt_bbox_PRED_gaze_PRED_act_PRED_weight_fusion_fine_token_only (Ex.1)
2. volleyball-dual-mid_p_p_field_middle_p_s_davt_bbox_GT_gaze_GT_act_GT_weight_fusion_fine_token_only (Ex.2)

• DAVT

python train.py yaml/volleyball/train_ours_p_p.yaml
python train.py yaml/volleyball/train_ours.yaml

The following folder contains the trained models.

1. volleyball-dual-mid_p_p_field_middle_p_s_davt_bbox_PRED_gaze_PRED_act_PRED_p_s_only (Ex.1)
2. volleyball-dual-volleyball-dual-mid_p_p_field_middle_p_s_davt_bbox_GT_gaze_GT_act_GT_p_s_only (Ex.2)

• ISA

python train.py yaml/volleyball/train_ours_isa.yaml

The following folder contains the trained models.

1. volleyball-dual-isa_bbox_PRED_gaze_PRED_act_PRED (Ex.1)
2. volleyball-dual-isa_bbox_GT_gaze_GT_act_GT (Ex.2)

• Ours

python train.py yaml/videocoatt/train_ours_p_p.yaml
python train.py yaml/videocoatt/train_ours.yaml

The following folder contains the trained models.

1. videocoatt-dual-p_p_field_deep_p_s_davt_scalar_weight_fix (Ex.1)
2. videocoatt-dual-p_p_field_deep_p_s_davt_scalar_weight_fix_token_only_GT (Ex.2)

• DAVT
  Trained model is published in here (https://github.com/ejcgt/attention-target-detection)

• ISA

python train.py yaml/videocoatt/train_ours_isa.yaml

The following folder contains the trained models.

1. videocoatt-isa_bbox_PRED_gaze_PRED (Ex.1)
2. videocoatt-isa_bbox_GT_gaze_GT (Ex.2)

• HGTD

python train.py yaml/videoattentiontarget/train_hgt.yaml

The following folder contains the trained models.

1. videocoatt-videoattentiontarget-hgt-high (Ex.1 and Ex.2)

You can choose the model which you would like to evaluate in the yaml files.

• Ours and DAVT

python eval_on_volleyball_ours.py yaml/volleyball/eval.yaml

• ISA

python eval_on_videocoatt_isa.py yaml/volleyball/eval.yaml

• Ours and DAVT

python eval_on_videocoatt_ours.py yaml/videocoatt/eval.yaml

• ISA

python eval_on_videocoatt_isa.py yaml/videocoatt/eval.yaml

• HGTD

python eval_on_videocoatt_hgt.py yaml/videocoatt/eval.yaml

You can choose the model which you would like to evaluate in the yaml files.

python demo_ours.py yaml/volleyball/demo.yaml

python demo_ours.py yaml/videocoatt/demo.yaml