International Conference on Computer Vision (ICCV) 2023
Pixel retrieval, often referred to as segmented instance retrieval, represents an emerging area within computer vision. The primary objective is to identify pixels that represent a specific query object within a database. This involves the machine's ability to recognize, localize, and segment the query object in database images in real-time. Essentially, it's a pixel-level and instance-level recognition task, as illustrated in the figure below:
Pixel retrieval can be viewed as an advanced form of image retrieval, providing insights into which specific pixels correlate with the query object. When juxtaposed with semantic and instance segmentation, pixel retrieval demands a nuanced recognition capability for targets of varying granularity. For a comprehensive understanding, please refer to our research paper.
Pixel retrieval is not just a theoretical concept; it has practical implications. Our user study indicates that pixel-level annotations during the retrieval process can notably enhance user experience. Engage with our user-study through this link.