YOLO for High Resolution Images without Retraining

In this paper, we consider the application of YOLO (You Only Look Once) to high-resolution images, i.e., so-called 4K, or 8K images. When those large images are applied to YOLO, in general, the original image will be resized before being processed by YOLO. In this case, the detection accuracy varies according to the size of the resized image. As an example, when a 4K image is resized to 320×320 pixels, YOLO would fail to detect objects that make up less than about 0.7% to 2% of the entire image. Or, it would fail to detect large objects that occupy about 2% to 10%, or more of the entire image, when 4K image is used without resizing. These values are confirmed in this paper through our experiments. Then, we propose a novel YOLO construction method that enables us to expand the area of the detectable objects when applied to high-resolution images. The advantage of the proposed method is that it does not require the retraining using high-resolution images. Instead, the proposed method uses images of multiple resolutions and combines the detection results. The effectiveness of the proposed method is confirmed by the experiments.