A 3D CNN time capsule for video recognition

What it does

This repository houses PyTorch implementations of three well-known 3D convolutional architectures—C3D, R3D, and R2Plus1D—for video action recognition. It covers the full training pipeline for the UCF101 and HMDB51 benchmarks, including a preprocessing step that extracts frames from raw AVI files. The author also provides training logs and accuracy curves for the C3D model.

The interesting bit

The real value is side-by-side access to multiple spatiotemporal architectures in one place, though the README makes clear that only C3D has actually been trained and released with pretrained weights so far. The author also links to a separate web-game project built on the C3D model, hinting the work is intended as a starting point for applications.

Key highlights

• Implements C3D, R2Plus1D, and R3D in a single PyTorch codebase.
• Includes frame-extraction preprocessing for UCF101 and HMDB51.
• Provides downloadable pretrained weights for C3D only.
• Publishes training/validation/test accuracy and loss curves for C3D experiments.
• Author maintains a related web demo game using the C3D model.

Caveats

• Targets PyTorch 0.4.0, so expect dependency friction on modern hardware.
• Only C3D has pretrained weights and reported experiments; R3D and R2Plus1D are implemented but not yet validated by the author.
• Dataset splits are generated with sklearn rather than using the official UCF101/HMDB51 partitions.

Verdict

Worth a look if you need a readable, self-contained reference for 3D video CNNs and can tolerate some archaeology. Look elsewhere if you need production-ready, modern PyTorch implementations of R(2+1)D or R3D.