Small object segmentation with fully convolutional network based on overlapping domain decomposition

Abstract

We propose a new segmentation algorithm based on deep learning. To segment ice hockey players, a fully convolutional network (FCN) is adopted and fine-tuned with our augmented training data. The original FCN has difficulty segmenting small-size objects. To solve this problem, our method divides an input image into four overlapping sub-images and each image is fed into the deep learning network. After obtaining segmentation results from all sub-images, we combine them into a single result. The segmentation results should be consistent over time in video. Thus, our method tracks segments over time and removes false positives that appear for brief periods. Mathematically, we show that our overlapping subdivision process can be interpreted as overlapping domain decomposition methods, which enable the FCN to regularize over consecutive sub-images in training time. Experimental results demonstrate that our method accurately segments ice hockey players when they appear small and when there exists severe background clutter. Our method shows real-time performance.