Dynamic-DETR: Dynamic Perception for RT -DETR

Object detection is a crucial branch of computer vision. RT-DETR, a recent advancement building upon DETR, has showcased superior performance over YOLO in real-time object detection. However, RT-DETR may struggle with complex inputs due to its limited dynamic perception. To address this, we propose Dynamic-DETR, an enhanced object detection algorithm characterized by its dynamic and flexible nature. In the backbone stage, we introduce the ideas of adaptive dilation and deformability to enable the model to adapt to complex scale variations and geometries; in the neck stage, we introduce content-aware feature reassembly to further enhance the dynamic perception of the model. Experiments have shown that our model outperforms RT -DETR in the 2D supervised object detection task on the KITTI and SODA10M datasets.