Back to projects▸DCGAN (Generator + Discriminator) implemented from scratch in PyTorch Lightning ▸Trained on HMDB51 (51-class human action recognition, ~7,000 clips) ▸Custom image sampler for Tensorboard visualisation during training ▸Benchmarked augmented clips against PyTorchVideo pre-trained classifiers
Research
DCGAN Video Augmentation
GAN-based data augmentation for human action recognition
Research implementation of a Deep Convolutional GAN for Human Action Recognition data augmentation on the HMDB51 dataset. Scalable training pipeline with multi-GPU support, benchmarked against pre-trained PyTorchVideo classifiers.
Tech stack
PyTorchPyTorch LightningPyTorchVideoPythonGANs
Problem
Medical imaging and action recognition datasets are often small and imbalanced. GANs offer a principled way to synthesise training data.
Implementation
Results
Generated synthetic video frames evaluated for fidelity via pre-trained classifier confidence scores. Training pipeline supports multi-GPU setups for scalable experimentation.