DogRecon: Canine Prior-Guided Animatable 3D Gaussian Dog Reconstruction From A Single Image

Abstract

We tackle animatable 3D dog reconstruction from a single image, noting the overlooked potential of animals. Particularly, we focus on dogs, emphasizing their intrinsic characteristics that complicate 3D observation. First, the considerable variation in shapes across breeds presents a complexity for modeling. Additionally, the nature of quadrupeds leads to frequent joint occlusions compared to humans. These challenges make 3D reconstruction from 2D observations difficult, and it becomes dramatically harder when constrained to a single image. To address these challenges, our insight is to combine the acquisition of appearance from generative models, without additional data, with geometric guidance provided by a parametric representation, aiming to achieve complete geometry. To this end, we present DogRecon, our framework consists of two key components: Canine-centric novel view synthesis with canine prior for multi-view generation of dog and a reliable sampling weight strategy with Gaussian Splatting for animatable 3D dog reconstruction. Extensive experiments on the GART, DFA, and internet-sourced datasets confirm our framework has state-of-the-art performance in image-to-3D generation and comparable performance in animatable 3D reconstruction. Additionally, we demonstrate novel pose animation and text-to-3D dog reconstruction as applications. Project page: https://vision3d-lab.github.io/dogrecon/