A COARSE-TO-FINE OBJECT DETECTION FRAMEWORK FOR HIGH-RESOLUTION IMAGES WITH SPARSE OBJECTS

To detect sparse small objects in high resolution images at a low cost is significantly more challenging than regular detection tasks. Compared to the overall detection accuracy, the recall rate is much less affected when using properly downsampled images for detection. Based on this fact, we propose a clustering-based coarse-to-fine object detection framework to enhance the object detection of sparse small objects. The first stage is coarse detection on a downsampled image to obtain image chips based on a clustering-baed region generation method. After that, the associated high resolution image clips are sent to a second-stage detector for fine detection. This approach reduces the number of chips for final object detection compared to regular methods, which divide the image into small tiles of the same size, and makes the best use of information in high-resolution images to increase detection accuracy. Experimental results show that our proposed approach achieves promising performance compared with other state-of-the-art detectors.