[ICCV2023] D3G:Exploring Gaussian Prior for Temporal Sentence Grounding with Glance Annotation
Datasets
- Annotations. We adopt the glance annotation released by ViGA for training. Specifically, we adapt the glance annotation to our framework in json format ( see in dataset ).
- Features. As for video features, we utilize the publicly available features for fair comparison following MMN.
Please download the video features to directory dataset as follows.
dataset
├── Charades_STA
│ ├── vgg_rgb_features.hdf5
│ ├── glance_charades_train.json
│ ├── charades_test.json
├── ActivityNet
│ ├── sub_activitynet_v1-3.c3d.hdf5
│ ├── glance_train.json
│ ├── val.json
│ ├── test.json
├── TACoS
│ ├── tall_c3d_features.hdf5
│ ├── glance_train.json
│ ├── val.json
│ ├── test.json
Main Results
Charades-STA Dataset
| Method |
Rank1@0.5 |
Rank1@0.7 |
Rank5@0.5 |
Rank5@0.7 |
| ViGA |
36.56 |
16.10 |
48.90 |
25.86 |
| D3G |
41.64 |
19.60 |
79.25 |
49.30 |
ActivityNet Captions Dataset
| Method |
Rank1@0.3 |
Rank1@0.5 |
Rank1@0.7 |
Rank5@0.3 |
Rank5@0.5 |
Rank5@0.7 |
| ViGA |
59.78 |
35.39 |
16.25 |
72.19 |
53.19 |
32.69 |
| D3G |
58.25 |
36.68 |
18.54 |
87.84 |
74.21 |
52.47 |
TACoS Dataset
| Method |
Rank1@0.3 |
Rank1@0.5 |
Rank1@0.7 |
Rank5@0.3 |
Rank5@0.5 |
Rank5@0.7 |
| ViGA |
20.82 |
9.52 |
3.10 |
27.92 |
15.35 |
6.10 |
| D3G |
26.99 |
12.62 |
4.77 |
54.71 |
31.59 |
12.10 |
Training & Inference
cd scipts
### charades
sh charades_train.sh # train
sh charades_test.sh # test
### activitynet
sh anet_train.sh # train
sh anet_test.sh # test
### tacos
sh tacos_train.sh # train
sh tacos_test.sh # test