Existing methods for instance segmentation in videos typi-cally involve multi-stage pipelines that follow the tracking-by-detectionparadigm and model a video clip as a sequence of images. Multiple net-works are used to detect objects in individual frames, and then associatethese detections over time. Hence, these methods are often non-end-to-end trainable and highly tailored to specific tasks. In this paper, we pro-pose a different approach that is well-suited to a variety of tasks involvinginstance segmentation in videos. In particular, we model a video clip asa single 3D spatio-temporal volume, and propose a novel approach thatsegments and tracks instances across space and time in a single stage. Ourproblem formulation is centered around the idea of spatio-temporal em-beddings which are trained to cluster pixels belonging to a specific objectinstance over an entire video clip. To this end, we introduce (i) novel mix-ing functions that enhance the feature representation of spatio-temporalembeddings, and (ii) a single-stage, proposal-free network that can rea-son about temporal context. Our network is trained end-to-end to learnspatio-temporal embeddings as well as parameters required to clusterthese embeddings, thus simplifying inference. Our method achieves state-of-the-art results across multiple datasets and tasks. Code and modelsare available at https://github.com/sabarim/STEm-Seg.
Download