DSINE

Why we included this project

DSINE is the official code behind a CVPR 2024 oral paper that rethinks how surface normal estimation models are built. Instead of treating normal prediction as a generic dense task, it feeds the model each pixel's ray direction and learns the relative rotation between neighboring normals, which yields crisp yet piecewise-smooth outputs on in-the-wild images of arbitrary resolution and aspect ratio. The pretrained weights, a minimal test script, and real-time demos for screen, webcam, and RealSense capture make it easy to gauge output quality before integrating it into a 3D reconstruction or SLAM pipeline. The authors also report stronger generalization than a recent ViT-based model despite training on a dataset orders of magnitude smaller, and the benchmark scripts covering six evaluation datasets are handy when comparing against other approaches.