Propagating Semantic Labels in Video Data

Semantic Segmentation combines two sub-tasks: the identification of pixel-level image masks and the application of semantic labels to those masks. Recently, so-called Foundation Models have been introduced; general models trained on very large datasets which can be specialized and applied to more specific tasks. One such model, the Segment Anything Model (SAM), performs image segmentation. Semantic segmentation systems such as CLIPSeg and MaskRCNN are trained on datasets of paired segments and semantic labels. Manual labeling of custom data, however, is time-consuming. This work presents a method for performing segmentation for objects in video. Once an object has been found in a frame of video, the segment can then be propagated to future frames; thus reducing manual annotation effort. The method works by combining SAM with Structure from Motion (SfM). The video input to the system is first reconstructed into 3D geometry using SfM. A frame of video is then segmented using SAM. Segments identified by SAM are then projected onto the the reconstructed 3D geometry. In subsequent video frames, the labeled 3D geometry is reprojected into the new perspective, allowing SAM to be invoked fewer times. System performance is evaluated, including the contributions of the SAM and SfM components. Performance is evaluated over three main metrics: computation time, mask IOU with manual labels, and the number of tracking losses. Results demonstrate that the system has substantial computation time improvements over human performance for tracking objects over video frames, but suffers in performance.