The Amazon Prime collaboration focused on the detection of artifacts and distortions present in streaming applications. Specifically, we aim to detect both transient and persistent artifacts within professional video feeds with long time horizons and no reference. Previous approaches to artifact detection lack the computational efficiency and generalizability to identify the wide range of potential distortions that may occur while monitoring thousands of streams at scale.
Our solution extracts features across disjoint domains combining standard natural scene statistics with novel ones including Laws’ filter responses. These features are then scored against a pristine reference distribution to characterize deviations from normal and pooled across frames in a short temporal window. Standard green learning tools, including the DFT and LNT, are utilized to identify and generate additional highly discriminative features for an XGBoost classifier. The green learning tools are organized into a multi-round framework, so that each successive round iteratively improves upon previous predictions for difficult samples. Experiments demonstrate that our method is both parameter efficient and accurate with performance exceeding comparable deep learning methods. In parallel, we developed a dataset of short and long form videos, based on LongVideoBench, with various streaming-specific distortions to aid further research in long-form professional video artifact detection.
