Low-light video processing has lagged behind image work partly because good ground-truth data is scarce. Existing datasets often limit motion, cover few scenes, use weak degradations, rely on extra sensors like event cameras, or suffer from misalignment between inputs and ground truth. They also tend to treat enhancement and scene understanding as separate problems, even though real applications need both.

To address this, the authors introduce OLiVES, a benchmark built from outdoor dynamic scenes captured with cinema-grade cameras. It includes over 25,000 aligned normal- and low-light frame pairs and more than 46,000 video object segmentation annotations, with content chosen to cover natural textures, animals, occlusions, dynamic activity, and close-ups.

The paper benchmarks current low-light video enhancement and video object segmentation methods on OLiVES, reporting that training on the dataset improves performance for both tasks. Because this is a single source, there are no conflicting findings to compare; the contribution is the dataset itself and the initial results it enables.