【摘要】 Transformer的自注意力机制允许模型处理长距离的依赖关系，从而能够有效处理复杂的空间关系并提供全局上下文信息，这使得基于Transformer的人体关键点检测展现出巨大的潜力。虽然Transformer在性能上表现出色，但其计算需求也相对较高。为此，模型的轻量化工作就尤为重要。文中实现了轻量化的Transformer进行人体关键点检测，在保持计算结果精度不变甚至有所提高的情况下，明显减少参数以及降低计算量。在COCO验证集上达到了75.1 AP的成绩。同时，相较于传统Transformer模型，其参数量降为1/2甚至1/3。更多还原

【Abstract】 The self-attention mechanism of the Transformer allows the model to process long-distance dependencies, thereby effectively handling complex spatial relationships and providing global contextual information. This enables human keypoint detection based on the Transformer to exhibit tremendous potential. Although the Transformer performs exceptionally in terms of performance, its computational requirements are relatively high. Therefore, the light-weighting of the model becomes particularly important. We have implemented a lightweight Transformer for human keypoint detection, which significantly reduces the number of parameters and computational load while maintaining or even improving the accuracy of the computation results. It achieved a score of 75.1 AP on the COCO validation set. At the same time, compared to mainstream Transformer models, the number of parameters of our model has been reduced to 1/2 or even 1/3.更多还原