pytorch-vsumm-reinforce

Why we included this project

This repo is the reference implementation for an AAAI 2018 paper on video summarization without labeled data. Instead of supervised learning, it trains a policy network with reinforcement learning, applying a diversity-representativeness reward so the chosen frames stay varied and representative of the source footage. The project includes the complete pipeline: splitting datasets like SumMe, training, evaluating against ground-truth summaries, plotting reward evolution, and converting the binary selection back into an mp4 file. The codebase is short enough to read from top to bottom, which makes it a good base for adapting the method to keyframe selection or highlight detection. One caveat: it targets PyTorch 0.4.0 and Python 2.7, so expect to port it to modern toolchains before running it.