Practical overview
When it fits
Use it when you can run a Python/PyTorch command-line workflow locally and need segment-level action classifications or video features.
Who it is for
- Developers building video-processing workflows that need action scores or features.
- Researchers exploring video action recognition with a pretrained 3D ResNet.
Documented uses
- Classify actions in video clips and inspect per-segment prediction scores.
- Generate 512-dimensional features from video segments for downstream analysis.
- Visualize classification results in generated videos.
Requirements & limits
The repository is public and MIT-licensed. Its documented use requires Python 3, PyTorch, FFmpeg, FFprobe, the code and a downloaded pretrained model.
- The project processes video in 16-frame segments, so its outputs are segment-level rather than a single clip-level result.
- Score and feature modes serve different downstream needs: class labels and scores versus 512-dimensional representations.
- The README names ResNeXt-101 as the best-performing option in the authors’ experiments; its command example also uses a resnet-34 model filename.
- Video handling depends on FFmpeg and FFprobe in addition to the Python and PyTorch stack.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Calculate action scores or video features
Run the documented command-line workflow to produce class scores or feature vectors for video segments.
Before you start
- Python 3 and PyTorch
- FFmpeg and FFprobe
- Downloaded project code and a pretrained model
- Input video files and an input list
Steps
Place video files in the video directory and prepare the input list.
Run main.py with the input list, video root, output path, model file and --mode score for class scores, or --mode feature for features.
For score output, use the documented result-video script to visualize classification results.
Expected outcome
Score mode produces action class names and prediction scores per 16 frames; feature mode produces 512-dimensional features per 16 frames.
Guide limitations
- The README does not specify the input-list format.
- The documented workflow does not detail the visualization script's arguments.
- The README's example model filename must correspond to the model file downloaded.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- kenshohara/video-classification-3d-cnn-pytorch · Official homepage title · retained in its source language