Skip to content

Latest commit

 

History

History
50 lines (46 loc) · 2.99 KB

README.md

File metadata and controls

50 lines (46 loc) · 2.99 KB

Prepare THUMOS14 Data

Get Preprocessed Data

We provide preprocessed data files to directly start training and evaluating.

  • The following files have been included in this repository: 
     data/thumos14/val/thumos14_val_annotation.csv
 data/thumos14/val/gt_info.pkl
 data/thumos14/val/gt_label.pkl
 data/thumos14/val/window_info.log
 data/thumos14/test/thumos14_test_annotation.csv
 data/thumos14/test/window_info.log
  • All relevant sliding window features of THUMOS14 are as follows. Download files from Onedrive according to your need. Extract them somewhere (refer to my directory structure in Onedrive or put them somewhere else and update config.py accordingly). For example, if you want to test temporal model pretrained on UCF101, you only need to download test/temporalDataXUCF101. The zip file is about 2.1G and the unzip folder is about 5.4G. There are 4741 npy files (named from 0.npy to 4740.npy) in each subfolder for validation set and 5103 npy files (named from 0.npy to 5102.npy) in each subfolder for test set.
    • test/spatialDataXUCF101 (2.5G -> 5.4G)
    • test/temporalDataXUCF101 (2.1G -> 5.4G)
    • test/spatialDataXKnetV3 (5.2G -> 10.7G)
    • test/temporalDataXKnetV3 (5.1G -> 10.7G)
    • val/spatialDataXUCF101 (2.4G -> 5.0G)
    • val/temporalDataXUCF101 (2.0G -> 5.0G)
    • val/spatialDataXKnetV3 (4.9G -> 10G)
    • val/temporalDataXKnetV3 (4.8G -> 10G)
  • (Optional) If you prefer to download data from Baidu Wangpan, here's the link and the key is K947. Due to the upload limitation of one file size, we compressed the whole directory of released data/models into a zip and them split them into 41 zip files (each is 700M, except that the last one is 360M). Please download them all and them extract them together.

Preprocess Data by Yourself

This part is not necessary and not recommended if you get preprocessed data above. If you want to start from the original THUMOS14 dataset and preprocess data by yourself, following are the instructions.

  1. Download THUMOS14 dataset (only validation and test set are needed)
  2. Extract frames and optical flow images refer to TSN.
  3. Generate validation and test set ground truth annotations refer to data/gen_thumos14_anno.py, which is based on SSN. 
    data/thumos14/test/thumos14_test_annotation.csv
data/thumos14/val/thumos14_val_annotation.csv
  4. These files related to data information will be generated by python data/get_data_info.py. 
    data/thumos14/val/gt_info.pkl
data/thumos14/val/gt_label.pkl
data/thumos14/val/window_info.log
data/thumos14/test/window_info.log
  5. Extract sliding window features refer to data/extract_feature.sh