Coarse-to-Fine Video Instance Segmentation With Factorized Conditional Appearance Flows

We introduce a novel method using a new generative model that automatically learns effective representations of the target and background appearance to detect, segment and track each instance in a video sequence. Differently from current discriminative tracking-by-detection solutions, our proposed hierarchical structural embedding learning can predict more high-quality masks with accurate boundary details over spatio-temporal space via the normalizing flows. We formulate the instance inference procedure as a hierarchical spatio-temporal embedded learning across time and space. Given the video clip, our method first coarsely locates pixels belonging to a particular instance with Gaussian distribution and then builds a novel mixing distribution to promote the instance boundary by fusing hierarchical appearance embedding information in a coarse-to-fine manner. For the mixing distribution, we utilize a factorization condition normalized flow fashion to estimate the distribution parameters to improve the segmentation performance. Comprehensive qualitative, quantitative, and ablation experiments are performed on three representative video instance segmentation benchmarks (i.e., YouTube-VIS
19
, YouTube-VIS
21
, and OVIS) and the effectiveness of the proposed method is demonstrated. More impressively, the superior performance of our model on an unsupervised video object segmentation dataset (i.e., DAVIS
19
) proves its generalizability. Our algorithm implementations are publicly available at https://github.com/zyqin19/HEVis.