Video Grounding

Dataset and Code

• Download dataset and code here

Structure of Video Grounding Dataset

• Annotations are stored in .txt format
• Annotations:
  • clip segment_start_time segment_end_time ## "annnotation_sentence"

Evaluation Metrics

• We follow VSLNet and use both Recall@1 and Mean IoU
• For the evaluation code, please refer to https://github.com/IsaacChanghau/VSLNet/blob/master/util/runner_utils_t7.py

Instructions to run Video Grounding models

1. To prepare the environment, please refer to
   • [LGI] https://github.com/JonghwanMun/LGI4temporalgrounding
   • [VSLNet] https://github.com/IsaacChanghau/VSLNet

2. For feature extraction, we used the I3D model and frame rate of 24 fps.

3. Move and replace files according to the directories in $DIR_AK_VG/video_grounding/code/code_new:
   • Helper script to move / create symbolic links to files
     • Remember to change the root directory $DIR_ROOT in $DIR_AK_VG/video_grounding/code/code_new/prepare_dir_VG.sh
     • bash $DIR_AK_VG/video_grounding/code/code_new/prepare_dir_VG.sh