Unsupervised Video Object Segmentation via Weak User Interaction and Temporal Modulation

In unsupervised video object segmentation (UVOS), the whole video might segment the wrong target due to the lack of initial prior information. Also, in semi-supervised video object segmentation (SVOS), the initial video frame with a fine-grained pixel-level mask is essential to good segmentation accuracy. It is expensive and laborious to provide the accurate pixel-level masks for each training sequence. To address this issue, We present a weak user interactive UVOS approach guided by a simple human-made rectangle annotation in the initial frame. We first interactively draw the region of interest by a rectangle, and then we leverage the mask RCNN (region-based convolutional neural networks) method to generate a set of coarse reference labels for subsequent mask propagations. To establish the temporal correspondence between the coherent frames, we further design two novel temporal modulation modules to enhance the target representations. We compute the earth mover's distance (EMD)-based similarity between coherent frames to mine the co-occurrent objects in the two images, which is used to modulate the target representation to highlight the foreground target. We design a cross-squeeze temporal modulation module to emphasize the co-occurrent features across frames, which further helps to enhance the foreground target representation. We augment the temporally modulated representations with the original representation and obtain the compositive spatio-temporal information, producing a more accurate video object segmentation (VOS) model. The experimental results on both UVOS and SVOS datasets including Davis2016, FBMS, Youtube-VOS, and Davis2017, show that our method yields favorable accuracy and complexity. The related code is available.