VIT中的Token一般是图像patch转换来的，所有token比重相同，本文提出基于聚类的Token，自适应调整Token的shape， size
- Clustering-based Token Merge (CTM) Block generates vision tokens of various locations, sizes, and shapes for each image by progressive clustering and merging tokens.
- Multi-stage Token Aggregation (MTA) head aggregates token features in multiple stages
- 在Transformer Block 中去掉了位置编码，使用 depth-wise 卷积层来获得位置特征
- 聚类算法
- Whole Body Pose Estimation