UniK3D

Why we included this project

UniK3D estimates metric 3D structure from a single RGB image without assuming a particular camera model, which is what sets it apart from typical monocular depth models. It accepts pinhole, fisheye, and panoramic inputs, so it stays useful in robotics, autonomous driving, and AR work where camera intrinsics change between setups. The repo includes pretrained ViT-S/B/L checkpoints you can load through Hugging Face or TorchHub, a Gradio demo for testing on your own images, and the full training and inference code. When you have known camera parameters or rays, the model can take those as input as well. The paper's zero-shot evaluation across different field-of-view categories gives teams a concrete sense of how the model generalizes before they commit to it.