Although Swin Transformer has made significant influence in computer vision, it does not perform well in object recognition and instance segmentation tasks due to model scale and complexity. Most existing approaches prefer addressing the above issues by altering model's architecture, which leads to additional pre-training costs. Therefore, we propose an efficient feature enhancement method and a residual attention mechanism. The feature enhancement method leverages CNN to augment the semantic features of Swin Transformer, while the residual attention mechanism prevents feature collapse. We evaluate our approach on the COCO 2017 val dataset, and the results demonstrate that our work achieves improved performance.
Zhang et al. (Tue,) studied this question.