Researchers have introduced MixSIS, a novel task for instance segmentation that allows for a flexible combination of annotation levels, including image, point, box, and mask annotations within a single dataset. This approach aims to optimize the annotation process by enabling detailed annotations for challenging objects while using less detailed annotations for simpler instances, thereby saving time and resources. To address this task, a baseline system called AnnotationsMixer (AnnsMixer) was developed, which generates pseudo mask annotations from all available annotation types, even when mixed within the same image. Evaluations show that AnnsMixer significantly outperforms existing methods on both weakly-supervised and MixSIS tasks, with case studies demonstrating substantial reductions in annotation time or improvements in results. AI
IMPACT Enables more efficient and flexible data annotation for computer vision tasks, potentially accelerating development in the field.
RANK_REASON The cluster describes a new research task and a baseline system presented in an arXiv paper. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
