we are hiring talented interns: [email protected]
In this paper, we study the problem of moment localization with natural language, and propose a novel 2D Temporal Adjacent Networks(2D-TAN) method. The core idea is to retrieve a moment on a two-dimensional temporal map, which considers adjacent moment candidates as the temporal context. 2D-TAN is capable of encoding adjacent temporal relation, while learning discriminative feature for matching video moments with referring expressions. Our model is simple in design and achieves competitive performance in comparison with the state-of-the-art methods on three benchmark datasets.
Please check the ms-2d-tan branch for our TPAMI extension.
- 🍻 Our journal extension is accepted by TPAMI.
- 🔧 A third-party optimized implementation by @ChenJoya.
- ☀️ Our paper was accepted by AAAI-2020. Arxiv Preprint
- 🏆 We extend our 2D-TAN approach to the temporal action localization task and win the 1st place in HACS Temporal Action Localization Challenge at ICCV 2019. For more details please refer to our technical report.
| Method | [email protected] | [email protected] | [email protected] | [email protected] |
|---|---|---|---|---|
| Pool | 40.94 | 22.85 | 83.84 | 50.35 |
| Conv | 42.80 | 23.25 | 80.54 | 54.14 |
I fixed a bug for loading charades visual features, the updated performance is listed above. Please use these results when comparing with our AAAI paper.
| Method | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] |
|---|---|---|---|---|---|---|
| Pool | 59.45 | 44.51 | 26.54 | 85.53 | 77.13 | 61.96 |
| Conv | 58.75 | 44.05 | 27.38 | 85.65 | 76.65 | 62.26 |
| Method | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] |
|---|---|---|---|---|---|---|
| Pool | 47.59 | 37.29 | 25.32 | 70.31 | 57.81 | 45.04 |
| Conv | 46.39 | 35.17 | 25.17 | 74.46 | 56.99 | 44.24 |
- pytorch 1.1.0
- python 3.7
- torchtext
- easydict
- terminaltables
Please download the visual features from box or dropbox and save it to the data/ folder.
Use the following commands for training:
# Evaluate "Pool" in Table 1
python moment_localization/train.py --cfg experiments/charades/2D-TAN-16x16-K5L8-pool.yaml --verbose
# Evaluate "Conv" in Table 1
python moment_localization/train.py --cfg experiments/charades/2D-TAN-16x16-K5L8-conv.yaml --verbose
# Evaluate "Pool" in Table 2
python moment_localization/train.py --cfg experiments/activitynet/2D-TAN-64x64-K9L4-pool.yaml --verbose
# Evaluate "Conv" in Table 2
python moment_localization/train.py --cfg experiments/activitynet/2D-TAN-64x64-K9L4-conv.yaml --verbose
# Evaluate "Pool" in Table 3
python moment_localization/train.py --cfg experiments/tacos/2D-TAN-128x128-K5L8-pool.yaml --verbose
# Evaluate "Conv" in Table 3
python moment_localization/train.py --cfg experiments/tacos/2D-TAN-128x128-K5L8-conv.yaml --verbose
Our trained model are provided in box or dropbox. Please download them to the checkpoints folder.
Then, run the following commands for evaluation:
# Evaluate "Pool" in Table 1
python moment_localization/test.py --cfg experiments/charades/2D-TAN-16x16-K5L8-pool.yaml --verbose --split test
# Evaluate "Conv" in Table 1
python moment_localization/test.py --cfg experiments/charades/2D-TAN-16x16-K5L8-conv.yaml --verbose --split test
# Evaluate "Pool" in Table 2
python moment_localization/test.py --cfg experiments/activitynet/2D-TAN-64x64-K9L4-pool.yaml --verbose --split test
# Evaluate "Conv" in Table 2
python moment_localization/test.py --cfg experiments/activitynet/2D-TAN-64x64-K9L4-conv.yaml --verbose --split test
# Evaluate "Pool" in Table 3
python moment_localization/test.py --cfg experiments/tacos/2D-TAN-128x128-K5L8-pool.yaml --verbose --split test
# Evaluate "Conv" in Table 3
python moment_localization/test.py --cfg experiments/tacos/2D-TAN-128x128-K5L8-conv.yaml --verbose --split test
If any part of our paper and code is helpful to your work, please generously cite with:
@InProceedings{2DTAN_2020_AAAI,
author = {Zhang, Songyang and Peng, Houwen and Fu, Jianlong and Luo, Jiebo},
title = {Learning 2D Temporal Adjacent Networks forMoment Localization with Natural Language},
booktitle = {AAAI},
year = {2020}
}