Home▸AI & Machine Learning▸Two-Stream Action Recognition: Spatial vs Temporal Fusion

Two-Stream Action Recognition: Spatial vs Temporal Fusion

Watch a procedural skeleton perform four actions while a spatial (pose-shape) stream and a temporal (motion-velocity) stream each vote on what it's doing, late-fused into one prediction — the two-stream architecture behind real video action recognition.

AI & Machine Learning3DModerate60 FPS📱 Mobile-adapted⇄ 2D version

Real video action-recognition systems — from the original two-stream network to I3D and SlowFast — classify what's happening by combining two very different signals: an appearance stream that reads a single frame's shape, and a motion stream that reads how things move between frames. This simulation drives a kinematic skeleton through four actions (walk, wave, jump, squat), extracts a live pose-shape descriptor and a live joint-velocity descriptor, scores each against per-action templates with cosine similarity and softmax, and late-fuses the two predictions with an adjustable weight. A noise slider shows why the motion stream is the more fragile of the two in practice, and the trailing ghost dots make visible the short temporal window any flow-based stream actually needs to see motion at all.

⚙ Under the hood

A kinematic skeleton performs four actions while a spatial (pose-shape) stream and a temporal (joint-velocity) stream each vote on what it's doing, late-fused into one prediction — the two-stream architecture behind real video action recognition.

action recognitionvideo analysistwo-stream networkoptical flowkinematicsAI

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install