Multi-scale pyramid convolution transformer for remote-sensing object detection

Reliable object detection in remote sensing imageries (RSIs) is an essential and challenging task in surface monitoring. However, RSIs are normally obtained from a high-altitude top-down perspective, causing intrinsic properties such as complex background, aspect ratio, color and scale variations. These properties restrict the domain transfer of sophisticated detector on nature images to RSIs, thereby deteriorating the desired detection performance. To address this issue, we propose a multi-scale pyramid convolution Transformer (MPCViT) that alleviates the limitations of ordinary visual Transformer. Specifically, we firstly employ an improved CNN to extract image features, generating initial feature pyramid. Then, bidirectional feature aggregation strategy is further used to improve feature representation capacity through feature enhancement and aggregation steps. To facilitate deep interaction of global dependencies and local details, dual-route encoding mechanism is constructed in each Transformer encoder. During inference stage, an iterative sparse keypoint sampling head is devised to enhance the detection accuracy. The competitive experimental results on NWPU VHR-10 and DIOR verify the efficacy of the proposed MPCViT.