dataset_preparation

Dataset Preparation

Processing Order

1. Extract features: ./script_video2feature.sh
2. Generate lists:
   • no official splits:
     1. ./script_dataset2split.sh
     2. ./script_dataset2list.sh
   • official splits exist (e.g. UCF, HMDB): ./script_list2DA.sh

Notes

1. The options in the scripts have comments with the following types:
   • no comment: user can still change it, but NOT recommend (may need to change the code or have different experimental results)
   • comments with choices (e.g. true | false): can only choose from choices
   • comments as depend on users: totally depend on users (mostly related to data path)

---

Feature vector (frame-level) extraction

video2feature.py loads the video dataset, and extract frame-level feature vectors, which are are needed for training and validation. A text file with all the category names is needed for labeled dataset.

Run ./script_video2feature.sh.

There are two modes:

• labeled: class_file=XXX/class_list_XXX.txt. The features with the categories only listed in the class_file will generated.
• unlabeled: class_file=none: all the features with the category "unlabeled" will be generated.

Output folder structures:

data_path/
  RGB-Feature/
    VIDEO_0001/
      img_00001.t7
      img_00002.t7
      ...
    VIDEO_0002/
    ...

---

Dataset split generation

dataset2split.py splits the whole dataset into two subsets according to the split ratio (randomly picking videos). Each subset has the same structure as shown above.

Run ./script_dataset2split.sh.

Options:

• input_type: depend on the format of raw data
• split_ratio: ratio of training data
• split_feat: if the features are already generated, users can split the corresponding features as well.

---

Data list generation

video_dataset2list.py generates the data list from a video dataset. The output text file will include all the selected categories showing {video_path, frame#, class}.

Run ./script_dataset2list.sh.

Options:

• random_each_video: If Y, it means users randomly select video clips from each raw and unsegmented video. It N, it means users randomly select video clips from each category.
• max_num: maximal numbers of selected video clips for each unsegmented video or category (depend on random_each_video)
• method_read: the method of calculating frame # of each video
  • video: load from the raw video folder (slower, but more accurate)
  • frame: load from the feature folder
• suffix: add some string to the list file name

Output text file (format: [video_full_path frame# class]):

data_path/RGB-Feature/VIDEO_0001/ 100 0
data_path/RGB-Feature/VIDEO_0002/ 150 1
......

---

DA list generation

If official splits exist (e.g. UCF, HMDB), list_ucf_hmdb_full2DA.py can separate the data list from the official splits with user chosen DA settings.

Run ./script_list2DA.sh.