Open World Instance Segmentation

Open-world instance segmentation (OWIS) aims to segment all objects within an image, including those unseen during training, unlike traditional closed-world methods. Current research focuses on developing models that leverage query-based approaches, transformers, and techniques like pseudo-annotation generation and contrastive learning to improve the detection and segmentation of novel objects. These advancements are crucial for applications requiring robust object recognition in dynamic environments, such as autonomous driving and robotics, where encountering unknown objects is inevitable. The field is actively exploring improved generalization capabilities and addressing challenges like handling highly overlapping predictions and mitigating biases towards known classes.