Table 1 Ablation study results on different subsets. Under the same training/test split and training strategy, squeeze-and-excitation (SE) attention mechanism, coordinate attention (CA), and dynamic confidence suppression (DCS) modules were progressively incorporated into the baseline HRNet-W48, and the corresponding performance changes were reported. In the table, “√/-” indicate enabled/disabled, respectively; all other settings follow the baseline. AP, Average precision; AP50, Average precision calculated at an object keypoint similarity (OKS) threshold of 0.50; AP75: Average precision calculated at an OKS threshold of 0.75; APM, Average precision for medium objects; APL, Average precision for large objects; AR, Average recall.

Table 2 Comparison of detection results of classical models on the Saihanwula wildlife dataset. All models adopt a top-down pose estimation framework and are evaluated on the test set. Evaluation follows the COCO Keypoint protocol. AP, Average precision; AP50, Average precision calculated at an object keypoint similarity (OKS) threshold of 0.50; AP75: Average precision calculated at an OKS threshold of 0.75; APM, Average precision for medium objects; APL, Average precision for large objects; AR, Average recall.

Table 3 Generalization test results based on the AP-10K dataset. We evaluate generalization on the AP-10K dataset, all methods are trained from scratch on the official AP-10K train/val split. AP, Average precision; AP50, Average precision calculated at an object keypoint similarity (OKS) threshold of 0.50; AP75: Average precision calculated at an OKS threshold of 0.75; APM, Average precision for medium objects; APL, Average precision for large objects; AR, Average recall.

Table 4 Experimental results of different methods on the self-built dataset and the AP-10K dataset (mean±SD).This table presents the average performance and standard deviation of various networks over five random seed experiments on two wildlife pose estimation datasets, reflecting the stability and variability of model performance. AP, Average precision; AP50, Average precision calculated at an object keypoint similarity (OKS) threshold of 0.50; AP75: Average precision calculated at an OKS threshold of 0.75; APM, Average precision for medium objects; APL, Average precision for large objects; AR, Average recall.

Table 5 Statistical significance and effect size analysis of different methods compared with HRNet on the self-constructed dataset and the AP-10K dataset. ∆AP denotes the difference in average precision (AP) compared with the baseline HRNet. The 95% CI represents the confidence interval (CI) of this difference, where a non-zero-crossing interval indicates statistical significance. The P-value assesses the statistical significance of the difference, with commonly used thresholds of P<0.05 (significant) and P<0.001 (highly significant). dn denotes the paired effect size (Cohen’s dn), which measures the magnitude of improvement, where 0.2/0.5/0.8 correspond to small/medium/large effects, respectively. Positive values indicate performance improvement, while negative values indicate performance degradation.

Table 6 A comparison of Giga floating-point operations per second (GFLOPs), parameter size (Params), and frames per second (FPS) across different methods. GFLOPs reflects the theoretical computational load of a single forward pass. Params indicates the number of trainable parameters, which is generally related to storage requirements and training difficulty. FPS denotes the average real-time processing frame rate under the test environment, gauging the model’s operational speed and deployment feasibility. Lower GFLOPs and parameter size typically indicate a lighter model, while a higher FPS suggests the model is more suitable for real-time application scenarios.