I3D (RGB + Flow)
The Inflated 3D (I3D) features are extracted using
a pre-trained model on Kinetics 400.
Here, the features are extracted from the second-to-the-last layer of I3D, before summing them up.
Therefore, it outputs two tensors with 1024-d features: for RGB and flow streams.
By default, it expects to input 64 RGB and flow frames (
224x224) which spans 2.56 seconds of the video recorded at 25 fps.
In the default case, the features will be of size
Tv x 1024 where
Tv = duration / 2.56.
Please note, this implementation uses either PWC-Net (the default)
and RAFT optical flow extraction instead of the TV-L1 algorithm,
which was used in the original I3D paper as it hampers speed.
Yet, it might possibly lead to worse peformance. Our tests show that the performance is reasonable.
You may test it yourself by providing
CUDA 11 and GPUs like RTX 3090 and newer
PWC optical flow back-end is not supported on CUDA 11 and, therefore, GPUs like RTX 3090 and newer.
RGB-only model should still work.
For details please check this issue #13
If you were able to fix it, please share your workaround.
Feel free to use
flow_type=raft RAFT during extraction.
The PWC-Net does NOT support using CPU currently
The PWC-Net uses
cupy module, which makes it difficult to turn to a version that does not use the GPU.
However, if you have solution, you may submit a PR.
||The number of frames from which to extract features (or window size).|
||The number of frames to step before extracting the next features.|
||I3D is a two-stream network. By default (
||By default, the flow-features of I3D will be calculated using optical from calculated with PWCNet (originally with TV-L1). Another supported model is
||If specified (e.g. as
||The device specification. It follows the PyTorch style. Use
||A list of videos for feature extraction. E.g.
||A path to a text file with video paths (one path per line). Hint: given a folder
||A path to a folder for storing the extracted features (if
||A path to a folder for storing temporal files (e.g. reencoded videos).|
Set up the Environment for I3D
Depending on whether you would like to use PWC-Net or RAFT for optical flow extraction,
you will need to install separate conda environments –
# it will create a new conda environment called 'pwc' on your machine conda env create -f conda_env_pwc.yml # or/and if you would like to extract optical flow with RAFT conda env create -f conda_env_torch_zoo.yml
Activate the environment
conda activate pwc
if you would like to use RAFT as optical flow extractor use
torch_zoo instead of
and extract features from
./sample/v_ZNVhz7ctTq0.mp4 video and show the predicted classes
python main.py \ feature_type=i3d \ device="cuda:0" \ video_paths="[./sample/v_ZNVhz7ctTq0.mp4]" \ show_pred=true
Activate the environment
conda activate pwc
The following will extract I3D features for sample videos. The features are going to be extracted with the default parameters.
python main.py \ feature_type=i3d \ device="cuda:0" \ video_paths="[./sample/v_ZNVhz7ctTq0.mp4, ./sample/v_GGSY1Qvo990.mp4]"
The video paths can be specified as a
.txt file with paths
python main.py \ feature_type=i3d \ device="cuda:0" \ file_with_video_paths=./sample/sample_video_paths.txt
It is also possible to extract features from either
flow modalities individually (
and, therefore, increasing the speed
python main.py \ feature_type=i3d \ streams=flow \ device="cuda:0" \ file_with_video_paths=./sample/sample_video_paths.txt
To extract optical flow frames using RAFT approach, specify
Note that using RAFT will make the extraction slower than with PWC-Net yet visual inspection of extracted flow
frames suggests that RAFT has a better quality of the estimated flow
# make sure to activate the correct environment (`torch_zoo`) # conda activate torch_zoo python main.py \ feature_type=i3d \ flow_type=raft \ device="cuda:0" \ file_with_video_paths=./sample/sample_video_paths.txt
The features can be saved as numpy arrays by specifying
--on_extraction save_numpy or
By default, it will create a folder
./output and will store features there
python main.py \ feature_type=i3d \ device="cuda:0" \ on_extraction=save_numpy \ file_with_video_paths=./sample/sample_video_paths.txt
You can change the output folder using
Also, you may want to try to change I3D window and step sizes
python main.py \ feature_type=i3d \ device="cuda:0" \ stack_size=24 \ step_size=24 \ file_with_video_paths=./sample/sample_video_paths.txt
By default, the frames are extracted according to the original fps of a video. If you would like to extract frames at a certain fps, specify
python main.py \ feature_type=i3d \ device="cuda:0" \ extraction_fps=25 \ stack_size=24 \ step_size=24 \ file_with_video_paths=./sample/sample_video_paths.txt
A fun note, the time span of the I3D features in the last example will match the time span of VGGish features with default parameters (24/25 = 0.96).
--keep_tmp_files is specified, it keeps them in
--tmp_path which is
./tmp by default.
Be careful with the
--keep_tmp_files argument when playing with
--extraction_fps as it may mess up the
frames you extracted before in the same folder.
- An implementation of PWC-Net in PyTorch
- The Official RAFT implementation (esp.
- A port of I3D weights from TensorFlow to PyTorch
- The I3D paper: Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset.
The wrapping code is MIT and the port of I3D weights from TensorFlow to PyTorch. However, PWC Net (default flow extractor) has GPL-3.0 and RAFT BSD 3-Clause.