kinetics-i3d

Why we included this project

If you need to recognize human actions in video clips, this repo is the reference implementation of the I3D architecture from the 2017 "Quo Vadis" paper, and it includes the actual pretrained weights, not just code. The two-stream model (RGB plus optical flow) was trained on Kinetics and gives strong action-recognition baselines right away; the bundled evaluation script runs a sample clip and prints the top predicted classes with probabilities. Researchers fine-tuning checkpoints on their own datasets, or teams exploring video classification for activity logging or content moderation, will find a solid starting point. It's a TensorFlow/Sonnet-era codebase from 2017, so treat it as a reproducible baseline to study and adapt rather than production software. The original preprocessing pipeline has since moved into MediaPipe, which the README points to if you need a maintained path for frame and flow extraction.