S3OD: Single Stage Small Object Detector from Scratch for Remote Sensing Images

Abstract

Small object detection is an important but challenge computer vision task in both natural scene and remote sensing scene. Due to the large difference of density, low contrast, sparse texture and arbitrary orientations, many advanced algorithms for small object detection in natural scene usually experience a sharp performance drop when directly applied to remote sensing images. In addition, most of state-of-the-art object detectors are fine-tuned from the off-the-shelf networks pretrained on large-scale classification dataset like ImageNet, which can incur learning bias and inconvenience of modification for remote sensing object detection tasks. In order to tackle these problems, a robust Single Stage Small Object Detector (S3OD) is trained from scratch, which can efficiently detect small-dense and small-dispersed objects in remote sensing images. The proposed S3OD adopts the small down-sampling factor to keep accurate location information and maintains high spatial resolution by introducing a new dilated residual block in deeper layers for small objects. Especially, the two-branch dilated feature attention module is proposed to enlarge the valid receptive field and make effective attention feature map for small-dense and small-dispersed object detection. S3OD can be trained from scratch stably while keeping the comparable performance by employing BatchNorm on both the backbone and detection head subnetworks. Experiments conducted on our built Remoting Sensing Small Object (RSSO) dataset shows that, our S3OD achieves the state-of-the-art accuracy for small objects detection and even performs better than several one-stage pretrained method.