Cascade multi-scale object detection on high-resolution images

Precise object detection is one of important task in computer vision. Recent achievements in convolutional neural networks open possibilities to detect objects with precision close to humans. However, current neural networks struggles to detect objects with large difference in scale. In this report proposed approach to process high-resolution images by same neural network multiple times with decreasing resolution in cascade way to enhance scale range of network pre-trained on typical dataset. Combined result of object detection of all passes processed by non-maximal suppression algorithm. Proposed approach demonstrated on Yolo3 network trained on COCO dataset. Scale range of network and upper size limit for detected object are estimated, scale technique for cascade decreasing resolution proposed.