video-classification-3d-cnn-pytorch

Why we included this project

Video understanding is a heavier lift than image classification, and this codebase handles the practical part: it ships pretrained 3D ResNet weights, plus the stronger ResNeXt variant, trained on the 400 action classes of Kinetics, so you can run inference on your own clips without training anything. Two modes cover the common workflows. Score mode labels each 16-frame window with a class name and confidence score, which is what you want for action recognition or video tagging. Feature mode extracts 512-dimensional embeddings after global average pooling, handy as input to a downstream classifier or retrieval pipeline. Researchers get a clean reference implementation of spatiotemporal CNNs as a bonus. One caveat: the code targets an older PyTorch/CUDA stack, so expect some adaptation work in a current environment.