Skip to content


The S3D action recognition model was originally introduced in Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification. We support the PyTorch weights for Kinetics 400 provided by According to the model card, with these weights, the model should achieve 72.08% top-1 accuracy (top5: 90.35%) on the Kinetics 400 validation set.

How the model was pre-trained? My best educated guess is that the model was trained on densely sampled 64-frame 224 x 224 stacks that were randomly trimmed and cropped from 25 fps 256 x 256 video clips (<= 10 sec). Therefore, to extract features (Tv x 1024), we resize the input video such that min(H, W) = 224 (?) and take the center crop to make it 224 x 224. By default, the feature extractor will split the input video into 64-stack frames (2.56 sec) with no overlap as it is during the pre-training and will do a forward pass on each of them. This should be similar to I3D behavior. For instance, given an ~18-second 25 fps video, the features will be of size 7 x 1024. Specify, step_size, extraction_fps, and stack_size to change the default behavior.

What is extracted exactly? The inputs to the classification head (see S3D.fc and S3D.forward) that were average-pooled across the time dimension.

Set up the Environment for S3D

Setup conda environment. Requirements are in file conda_env_torch_zoo.yml

# it will create a new conda environment called 'torch_zoo' on your machine
conda env create -f conda_env_torch_zoo.yml

Quick Start

Open In Colab

Activate the environment

conda activate torch_zoo

and extract features from the ./sample/v_GGSY1Qvo990.mp4 video and show the predicted classes

python \
    feature_type=s3d \
    video_paths="[./sample/v_GGSY1Qvo990.mp4]" \

See the config file for ther supported parameters.

Supported Arguments

stack_size 64 The number of frames from which to extract features (or window size). If omitted, it will respect the config of model_name during training.
step_size 64 The number of frames to step before extracting the next features. If omitted, it will respect the config of model_name during training.
extraction_fps 25 If specified (e.g. as 5), the video will be re-encoded to the extraction_fps fps. Leave unspecified or null to skip re-encoding.
device "cuda:0" The device specification. It follows the PyTorch style. Use "cuda:3" for the 4th GPU on the machine or "cpu" for CPU-only.
video_paths null A list of videos for feature extraction. E.g. "[./sample/v_ZNVhz7ctTq0.mp4, ./sample/v_GGSY1Qvo990.mp4]" or just one path "./sample/v_GGSY1Qvo990.mp4".
file_with_video_paths null A path to a text file with video paths (one path per line). Hint: given a folder ./dataset with .mp4 files one could use: find ./dataset -name "*mp4" > ./video_paths.txt.
on_extraction print If print, the features are printed to the terminal. If save_numpy or save_pickle, the features are saved to either .npy file or .pkl.
output_path "./output" A path to a folder for storing the extracted features (if on_extraction is either save_numpy or save_pickle).
keep_tmp_files false If true, the reencoded videos will be kept in tmp_path.
tmp_path "./tmp" A path to a folder for storing temporal files (e.g. reencoded videos).
show_pred false If true, the script will print the predictions of the model on a down-stream task. It is useful for debugging.


  1. The kylemin/S3D implementation.
  2. The S3D paper: Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification.