Publication

Abstract : We propose CANpose, a lightweight model for human pose recognition that integrates convolutional networks with a bi-modal cross-attention mechanism to effectively capture both local and global features. In contrast to conventional CNN-based and hybrid attention approaches, CANpose leverages the complementary nature of pose depth maps and binary silhouettes through cross-attention, enhancing its ability to distinguish between visually similar poses. By employing depth-wise separable convolutions for efficient local feature extraction, CANpose significantly reduces computational complexity–achieving up to 17× fewer parameters and 11× lower FLOPs–while maintaining high recognition accuracy across multiple benchmark datasets. These results position CANpose as a scalable and efficient solution for real-world pose recognition applications. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.