APG-DynaFFNet: Anatomical priors guided dynamic multi-granularity feature fusion for Animal Pose Estimation

Animal Pose est:imation is of great importance for behavioral analysis, individual identification, and ecological conservation. However, pose estimation in complex natural environments remains challenging due to severe occlusions, diverse postures, and high movement flexibility across species. Inspired by adaptive computing architectures and structured sparse representation learning, we propose APG-DynaFFNet, a novel animal pose estimation framework that explicitly integrates species-aware anatomical priors with dynamic multi-granularity feature fusion. Unlike conventional single-stage or fixed-fusion architectures, this framework introduces an adaptive two-stage inference strategy with quality-gated decision-making. In the first stage, it performs coarse keypoint localization while simultaneously assessing prediction quality and capturing spatial topological relationships among keypoints. In the second stage, it constructs sparse high-resolution representations to enhance fine-grained local features.Furthermore, a bidirectional reasoning module is developed to capture semantic relationships among keypoints, enabling structure-consistent pose predictions across diverse species. The proposed method can better overcome the challenges of computational redundancy and occlusion sensitivity through structured sparse processing and a structure-aware mechanism. Extensive experiments on a self-constructed multi-species animal pose dataset (including golden monkeys, lynxes, snow leopards, etc.) demonstrate the superior performance of APG-DynaFFNet. The model achieves an AP of 86.8% on the Golden Monkey dataset, outperforming multiple strong baselines. The full model further attains 87.7% AP on lynxes and snow leopards, confirming consistent gains across species. These results highlight the superior cross-species generalization and robustness of the proposed method, establishing it as a competitive solution for real-world animal pose estimation.