LLVIP

Why we included this project

LLVIP pairs visible-light and infrared shots of the same night-time street scenes, so it is a natural fit for anyone building low-light vision systems. Its roughly 30,000 images carry pedestrian bounding-box annotations, which means teams working on night surveillance or autonomous perception can train and benchmark object detectors on it right away. Because every scene appears in both spectra, the data also lends itself to image fusion and image-to-image translation work, and the repo ships baseline models like FusionGAN and Densefuse plus a toolbox that converts annotations between YOLO and COCO formats. Raw unregistered pairs and videos are available too, which extends its use to image registration research. This is a dataset and research companion, not a deployable application, so expect to use it for training data and comparison baselines rather than dropping it into production.