PySlowFast
title: PySlowFast type: language created: 2026-06-22 last_updated: 2026-06-22 related: ["radar/languages/Ultralytics", "radar/languages/DeepFace", "radar/techniques/DetectionThenIdentityVideoPipeline"] sources: ["https://github.com/facebookresearch/SlowFast", "https://blog.ippon.fr/2026/05/18/computer-vision-et-rugby-retour-dexperience-sur-la-construction-dun-pipeline-danalyse-video-avec-yolo-et-deepface/"] radar_quadrant: Languages & Frameworks radar_ring: Assess radar_position: outer
PySlowFast is a video understanding research codebase from Facebook AI Research (FAIR) implementing a family of temporal action recognition models. Supported architectures include SlowFast, Slow, C2D, I3D, Non-local Network, X3D, MViTv1, MViTv2, Rev-ViT, Rev-MViT, and self-supervised methods (MAE for Video, MaskFeat, contrastive SSL). Apache 2.0 license, 7.4k stars, 1.3k forks.
PySlowFast models classify video clips into action categories by modeling temporal dynamics across frames. SlowFast specifically uses two pathways: a slow pathway capturing spatial semantics at low frame rate and a fast pathway capturing motion at high frame rate. This makes it well-suited for distinguishing dynamic events (tackles, passes, jumps) that a static object detector cannot differentiate.
In a radar/techniques/DetectionThenIdentityVideoPipeline, PySlowFast fills a third stage after detection and identity: action classification within detected bounding boxes. Given a tracked subject (identified by radar/languages/Ultralytics and radar/languages/DeepFace), PySlowFast can label the action being performed per clip segment.
The repository is explicitly framed as "designed in order to support rapid implementation and evaluation of novel video research ideas." No versioned releases exist. Dataset setup requires separate documentation per model family. 424 open issues and 21 open pull requests indicate a backlog that a small maintainer team has not cleared. The rugby pipeline blog (Ippon Technologies, May 2026) attempted PySlowFast for dynamic action recognition (tackles, passes) but paused due to annotation data requirements: "travail colossal que nous avons dû mettre en pause."
The annotation burden is the primary practical barrier. Meaningful action classification requires a labeled training set specific to the domain; COCO-pretrained weights do not transfer to domain-specific actions the way general object detection does.
Placed in Languages & Frameworks / Assess / outer.
PySlowFast is the most established open-source temporal action recognition codebase and the natural next stage after object detection and face identity in a video analysis pipeline. The outer Assess position reflects the gap between research utility and pipeline composability: no versioned releases, no documented export format compatible with downstream editing tools, high annotation cost for custom domains, and confirmed abandonment in at least one production attempt. Worth evaluating for workflows where action classification is essential and a labeled domain dataset exists; otherwise the annotation investment is prohibitive relative to alternatives.
