Video Feature Extractor

January 17, 2020 ยท View on GitHub

This repo is for extracting video features.

Preparation

You just need make csv files which include video paths information. Please run python utils/build_dataset.py. See utils/build_dataset.py for more details.

Requirements

  • python 3.x
  • pytorch >= 1.0
  • torchvision
  • pandas
  • numpy
  • Pillow
  • h5py
  • tqdm
  • PyYAML
  • addict

Pretrained Models

Supported models are 3DResNet, SlowFastNetwork with non local block, (I3D).
You can download from here Pretrained I3D model is not available yet.

Extracting and Save Video Features

Please run python extract.py [dataset_dir] [save_dir] [csv] [arch] [pretrained_weights] [--sliding_window] [--size] [--window_size] [--n_classes] [--num_workers] [--temp_downsamp_rate [--file_format].

References