This is my implementation of BMN: Boundary-Matching Network for Temporal Action Proposal Generation(2019 ICCV) in pure Pytorch. It contains the full pipeline of training and evaluation on ActivityNet-1.3. The key features of this repo are:
- Full training and evaluation pipeline.
- Automatic parameter saving, loading and changing through command line.
Python version: 2 or 3
Packages:
- Pytorch >= 1.0
- pandas
- tqdm
After 9 epochs, we can get accuracy as below:
| AN | Recall |
|---|---|
| AR@1 | 33.6% |
| AR@5 | 49.7% |
| AR@10 | 57.0% |
| AR@100 | 75.4% |
| AUC | 67.4% |
All data is saved in folder 'data', specific list is as follows:
data
-- activitynet_annotations
-- · action_name.csv
-- · anet_anno_action.json
-- · video_info_new.csv
-- activitynet_feature_cuhk
-- · csv_mean_100
-- eval
-- · activity_net_1_3_new.json
-- · activity_net.v1-3.min.json
-- · sample_classification_prediction.json
-- · sample_detection_prediction.json
We use same extracted-feature and data as BSN.
We support experiments with publicly available dataset ActivityNet 1.3 for temporal action proposal generation now. To download this dataset, please use official ActivityNet downloader to download videos from the YouTube.
To extract visual feature, we adopt TSN model pretrained on the training set of ActivityNet, which is the challenge solution of CUHKÐ&SIAT team in ActivityNet challenge 2016. Please refer this repo TSN-yjxiong to extract frames and optical flow and refer this repo anet2016-cuhk to find pretrained TSN model.
For convenience of training and testing, we rescale the feature length of all videos to same length 100, and we provide the rescaled feature at here Google Cloud or Baidu Yun. If you download features from BaiduYun, please use cat zip_csv_mean_100.z* > csv_mean_100.zip before unzip. After download and unzip, please put csv_mean_100 directory to ./data/activitynet_feature_cuhk/ .
Using train.py for training. Check the opt.py for more details. You may set the parameters yourself in your own specific task.
python train.py -- epochs 30Same as above, using test.py for testing.
python test.pyI referred to many fantastic repos during the implementation:
Tianwei Lin-BMN-Paddle
JJBOY-BMN-Pytorch
[ ] mAP calculating.