ID-Pose: Sparse-view Camera Pose Estimation by Inverting Diffusion Models

• ArXiv: https://arxiv.org/abs/2306.17140
• Code: https://github.com/xt4d/id-pose/
• Demo: https://huggingface.co/spaces/tokenid/ID-Pose/

Abstract

• ID-Pose estimates camera poses of sparse-view images of a 3D object (appearance overlaps not required).
• ID-Pose inversely uses the off-the-shelf Zero123 to estimate camera poses by iteratively minimizing denoising errors given input images.
• ID-Pose is a zero-shot method that requires NO additional model training or finetuning.
• ID-Pose exhibits strong generalization ability on open-world images as the method effectively leverages the image priors from Zero123 (StableDiffusion).