AW-HRNet: A Spatially-Aware and Context-Enhanced Network for Few-Shot Animal Keypoint Detection

Animal keypoint detection is crucial for recognizing animal behavior. Current few-shot keypoint detection models perform well in cross-species tasks with novel keypoints; however, there exists insufficient modeling of topological relationships between keypoints, and excessive reliance on local features while neglecting global context are two major limitations. To solve the above challenges, the proposed AW-HRNet is designed to incorporate spatial awareness and contextual enhancement. The extraction of local fine-grained features is achieved through Adaptive Rectangular Convolution (ARConv), while global context is modeled by combining it with Wavelet Transform Convolution (WTConv), enabling joint optimization of local representation and global structure to be accomplished. In addition, AW-HRNet is applied to a few-shot keypoint detection network for animal keypoint detection. Compared to previous SOTA methods, our model achieves a 4.73% improvement on the Animal Pose dataset.