Video Object Segmentation Based on Location and RoIAlign in Weight Modulated Multi-Scale Network

As a binary labeling problem, video object segmentation targets at segmenting one or more specific objects throughout a video sequence. Since 2016, many video object segmentation methods based on deep learning have been developed. Most of these methods use the ground truth segmentation mask of the first frame to fine-tune the subsequent segmentation frames of the video, so that the deep network can remember the appearance of the objects that need to be segmented in a given video. Satisfactory segmentation accuracy can be achieved ultimately. But the fine-tuning process fails to meet the requirements of real-time applications. In this work, we propose a fast and effective method which does not rely on fine-tuning. Our network is mainly divided into two parts: location and segmentation. We also extract the semantic information of the annotation object in the first frame to generate corresponding channel-wise weights so as to re-target the network to locate and segment the specific object accurately. The experimental results show that on the DAVIS 2016 dataset, our method without fine-tuning exhibits more competitive results than the most advanced methods using online fine-tuning. The J&F average index reached 79.7%, and the running time per frame was only 0.11s.