Closed zkdfbb closed 2 years ago
The settings of heatmap-based methods discussed in our paper simply follow their original papers. Which means, for HRNet&SimBa, shift strategy is adopted as post-processing; for TokenPose, DARK strategy is adopted as post-processing. By contrast, none of them is used by our proposed method (no need of extra post-processing for further performance improvement).
By far, SimDR works in top-down human pose estimation. For bottom-up settings or else, we regard them as future work.
想请教一下,你论文里面的heatmap后处理都用了哪些方式,在小输入尺寸下差别还蛮大的
SimDR这方法可以用在多人姿态估计里,想了一下one-stage的话貌似不太方便