Geometry-Aware Animal Reconstruction via Depth Refinement and Adaptive Distillation

RGB-D-based animal reconstruction offers richer geometric cues than RGB-only approaches, enabling higher-precision 3D recovery and thus playing a crucial role in real-world applications. Due to the scarcity of annotated real-world data, models are compelled to rely heavily on large-scale synthetic datasets, yet the synthetic-to-real domain gap hinders progress, leading to significant degradation when applied to real captures. To address this issue, we propose GeoAni, a geometry-aware framework for robust 3D animal reconstruction from RGB-D inputs, whose core insight is to combine depth-refinement pretraining and geometry-guided adaptive distillation. GeoAni comprises three stages: (i) depth-refinement pretraining for geometry-aware RGB-D representations; (ii) synthetic supervised reconstruction of pose and shape; (iii) real-domain adaptation via geometry-guided distillation. Extensive experiments demonstrate that our framework achieves state-of-the-art robust and accurate animal reconstruction under diverse scenarios, effectively overcoming domain discrepancies and maintaining high precision under challenging conditions.