Referred Segmentation on Single/ No Target Image

This paper presents the 'Single or No' model, an innovative approach to referred segmentation that overcomes the limitations of conventional methods, which only focus on single-object segmentation while ignoring no-target scenarios. Leveraging the robust framework of Detectron2 and incorporating a fine-tuned CLIP model, the proposed model adeptly extracts objects and computes the similarity between text and image to effectively handle both single-target and no-target cases. Through comprehensive evaluations on the refCOCO and grefCOCO datasets, the model demonstrates significant improvements in segmentation accuracy. Furthermore, it provides valuable insights and potential applications for related tasks such as image description generation and visual question answering, paving the way for more versatile and accurate segmentation solutions.