DSINE
DSINE is a state-of-the-art surface normal estimation framework presented as an oral paper at CVPR 2024. Developed by Gwangbin Bae and Andrew J. Davison, it rethinks inductive biases for this specific task by moving away from general-purpose dense prediction models. The method uniquely utilizes per-pixel ray directions and learns the relative rotation between neighboring surface normals to encode their relationships. This approach enables the generation of crisp yet piecewise smooth predictions for challenging in-the-wild images of arbitrary resolution and aspect ratio. DSINE demonstrates strong generalization capabilities, outperforming recent Vision Transformer-based models despite being trained on an orders-of-magnitude smaller dataset. The software includes an official implementation with ready-to-use pre-trained weights for minimal setup testing or full benchmark evaluation. It supports standard hardware requirements via PyTorch and allows users to estimate uncertainty in predictions. The package facilit