Title: PRIMA: Enhancing Animal Mesh Recovery via Biological Priors and Test-Time Adaptation

Abstract: This paper introduces PRIMA (PRIors for Mesh Adaptation), a novel framework designed to achieve robust 3D quadruped mesh reconstruction even in the presence of significant imbalances in species representation and pose distribution. Current animal reconstruction techniques frequently suffer from poor generalization to rare articulations and underrepresented animals, as they tend to regress toward average shapes and poses. This limitation stems from insufficient 3D supervision and the long-tailed nature of species distributions. To overcome these hurdles, PRIMA offers three primary innovations. Firstly, the framework integrates BioCLIP embeddings to serve as biological priors, thereby injecting semantic and morphological insights into the reconstruction pipeline to facilitate more accurate and broadly applicable shape predictions for various quadrupeds. Secondly, we propose a test-time adaptation (TTA) mechanism that enhances SMAL predictions by applying 2D reprojection constraints and auxiliary keypoint guidance. This approach not only refines pose and shape estimation but also facilitates the creation of high-quality pseudo-3D annotations from existing 2D datasets. Thirdly, utilizing this TTA framework, we have developed Quadruped3D, a comprehensive large-scale dataset encompassing diverse species and pose variations, which systematically boosts model performance. Extensive evaluations conducted on the Animal3D, CtrlAni3D, Quadruped2D, and Animal Kingdom benchmarks reveal that PRIMA delivers state-of-the-art performance, with notable gains in handling underrepresented species and complex poses. These findings underscore the critical role of biological priors and adaptation-based data expansion in achieving scalable and generalizable animal mesh recovery. The source code is accessible at https://github.com/AdaptiveMotorControlLab/PRIMA.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC