This repo is an implementation of Efficient Video Generation on Complex Datasets
| Package | version |
|---|---|
| python | >=3.5 |
| pytorch | 1.12 |
| numpy | 1.17.2 |
| pandas | 0.25.1 |
| tensorboardX | 1.8 |
| ffmpeg | 3.4.2 |
Note: For more detail, please look up requirements.txt
sudo apt install ffmpeg # important package
chmod u+x scripts/data_prepare.sh
scripts/data_prepare.sh <dataset_path>
scripts/train_model.sh <runing_mode> <dataset_path>
Process UCF-101
- Step 1: Download dataset
- Step 2: Convert from avi to jpg files using: python utils/video_jpg_ucf101_hmdb51.py avi_video_directory jpg_video_directory
- Step 3: Generate n_frames files using: python utils/n_frames_ucf101_hmdb51.py jpg_video_directory
- Step 4: Generate annotation file in json format similar to ActivityNet using: python utils/ucf101_json.py annotation_dir_path
Note: To change the number of class:
- Modify classInd.txt to contain the expected class(es). For example: 1 ApplyEyeMakeup 2 ApplyLipstick 3 Archery
- Run step 4 only
- The code in dataloader automatically skips the unsed videos.