Unsupervised Video Object Segmentation via Weak User Interaction and Temporal Modulation

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