This is the repository for our ICCV 2017 paper TALL: Temporal Activity Localization via Language Query.
Download the C3D features for training set and test set of TACoS dataset. Modify the path to feature folders in main.py
Download the Skip-thought sentence embeddings and sample files from here of TACoS Dataset, and put them under exp_data folder.
python main.py
The sentence temporal annotations on Charades dataset are available here: train, test. The format is "[video name] [start time] [end time]##[sentence]". You may want to generate the skip-thought embeddings and C3D features on Charades-STA, and modify the codes slightly to reproduce the experiments.
I did some anno cleaning for Charades-STA (compared to the version I used in ICCV paper), the updated performance is listed below. Please compare to these results when using Charades-STA.
| Model | R@1,IoU=0.5 | R@1,IoU=0.7 | R@5,IoU=0.5 | R@5,IoU=0.7 |
|---|---|---|---|---|
| CTRL (aln) | 17.69 | 5.91 | 55.54 | 23.79 |
| CTRL (reg-p) | 19.22 | 6.64 | 57.98 | 25.22 |
| CTRL (reg-np) | 21.42 | 7.15 | 59.11 | 26.91 |