What it does

This repo implements five classic architectures for classifying human actions in video, using the UCF101 dataset of 101 action categories. You get frame-by-frame CNN classification, CNN-to-LSTM pipelines (both two-stage and end-to-end LRCN), CNN-to-MLP, and 3D convolutional networks. It’s essentially a working notebook made public: run the data extraction scripts, wait eight hours for feature extraction on a mid-tier GPU, then train your pick of models.

The interesting bit

The value is in the side-by-side comparison, not novelty. The author wired up five approaches that were standard circa 2017 so you can see how they differ in complexity and plumbing — from “just pretend video is a stack of images” to “actually model spatiotemporal cubes with 3D convolutions.” The LRCN variant (time-distributed CNN feeding an RNN in one network) is the most architecturally elegant of the bunch.

Key highlights

• Five model architectures defined in a single models.py for easy comparison
• Full data pipeline from raw UCF101 videos to frame sequences and CSV manifests
• Feature extraction cached to disk so LSTM/MLP training doesn’t re-run CNN forward passes
• TensorBoard and CSV logging built in
• Multiple worker support in the data generator (checked off the TODO list)

Caveats

• No demo script: you cannot point a finished model at a new video and get a prediction without writing code yourself
• Locked to Keras 2 and TensorFlow 1.x — this is legacy stack territory now
• Requires ffmpeg and manual path tweaking on non-Unix systems
• Data augmentation and optical flow are on the TODO list, not implemented

Verdict

Worth a look if you’re teaching computer vision or need a baseline to beat on UCF101. Skip it if you want production-ready video understanding — modern transformers and pre-trained video backbones have left this approach behind, and the dependency stack shows its age.