Learning Structure-Supporting Dependencies via Keypoint Interactive Transformer for General Mammal Pose Estimation

TL;DR

The paper tackles general mammal pose estimation across diverse species, a setting plagued by large appearance and pose variations and limited data per species. It introduces KITPose, a Keypoint-Interactive Transformer that learns instance-level structure-supporting dependencies by combining a high-resolution HRNet backbone, Generalised Heatmap Regression Loss for intermediate supervision, body part prompts derived from keypoint clustering, and an adaptive keypoint weighting scheme. The core contributions are the closed-loop integration of keypoint tokens with body-part prompts in a single-head attention KIT block, plus principled clustering-based body-part priors and adaptive losses that improve data efficiency. Empirical results on AP10K, AnimalKingdom, and COCO demonstrate state-of-the-art or competitive performance, with ablations validating the importance of each component and showing strong generalisation to human pose estimation. The work offers a practical, data-efficient approach to universal animal pose estimation and a pathway toward cross-species pose understanding in real-world applications.

Abstract

General mammal pose estimation is an important and challenging task in computer vision, which is essential for understanding mammal behaviour in real-world applications. However, existing studies are at their preliminary research stage, which focus on addressing the problem for only a few specific mammal species. In principle, from specific to general mammal pose estimation, the biggest issue is how to address the huge appearance and pose variances for different species. We argue that given appearance context, instance-level prior and the structural relation among keypoints can serve as complementary evidence. To this end, we propose a Keypoint Interactive Transformer (KIT) to learn instance-level structure-supporting dependencies for general mammal pose estimation. Specifically, our KITPose consists of two coupled components. The first component is to extract keypoint features and generate body part prompts. The features are supervised by a dedicated generalised heatmap regression loss (GHRL). Instead of introducing external visual/text prompts, we devise keypoints clustering to generate body part biases, aligning them with image context to generate corresponding instance-level prompts. Second, we propose a novel interactive transformer that takes feature slices as input tokens without performing spatial splitting. In addition, to enhance the capability of the KIT model, we design an adaptive weight strategy to address the imbalance issue among different keypoints.

Figures (9)

• Figure 1: (a) The obtained feature maps from the backbone and the proposed KIT module. (b) The visualisation of the attention maps of the KIT module.
• Figure 2: Overview of KITPose. Firstly, the instance features are extracted by the backbone. In the keypoint branch, the GHRL loss function is used as an intermediate supervision. Then the features are directly flattened to 1D tokens along the channel dimension. In the context branch, the NanoBlock is utilised to extract instance-level context tokens. The context tokens are then fused with the body part biases obtained through clustering of keypoint tokens to generate body part prompts. The structural relationship and constraint are captured through self-attention interactions in each Transformer block. Finally, an output head is used to predict the keypoints heatmaps.
• Figure 3: In the vision transformer process, spatial multi-head self-attention works on the split spatial feature patches, where each patch represents a spatial token. Each token is then divided into multiple heads along the channel dimension. In our design, channel single-head self-attention is explicitly performed in keypoint features. Attention takes the entire feature slice as a token. Relevant keypoints can be involved in the prediction of each keypoint to deliver supporting clues.
• Figure 4: Visualisation of the KITPose-E2C4 body part prompts and the interaction patterns. The leftmost column corresponds to the input instances, the middle 4 columns enclosed by blue dashed lines delineate the body part representations, and the rightmost column contains the attention map for information interactions. 'BP' and 'KPT' mean body part and keypoint respectively. The attention map can be divided into four parts: the orange box signifies BP-to-BP responses, the cyan box represents BP-to-KPT responses, the blue box denotes KPT-to-BP responses, and the red box indicates KPT-to-KPT responses.
• Figure 5: Performance sensitivity analysis in terms of the number of body part clusters $N_{p}$. The results are reported on AP10K and AnimalKingdom using the KITPose-E2 model, with an input size of $256\times256$.