Visionwhich

Visionwhich is a proposed framework in computer vision and human–computer interaction for resolving referential ambiguity in visual scenes by interpreting explicit which prompts and contextual cues to identify the intended target among several candidates. It describes a family of models that aim to map user intent expressed in natural language questions or prompts to a specific object or region within an image or video.

The term has appeared in early-2020s academic discussions and demonstration materials as part of broader work on interactive visual question answering and referential expression understanding. It is not a universally standardized term, but rather a descriptive label for methods that combine visual analysis with user-guided disambiguation to select a single target from multiple possibilities.

Techniques associated with visionwhich typically fuse visual encoders and language models, sometimes with dialog history or user feedback, to compute a probability distribution over candidate objects or regions. They leverage attributes, spatial relations, and context to answer questions such as “which cup did you mean?” and to return the most likely target along with a confidence score.

Applications span collaborative robotics, accessibility tools that assist users in cluttered environments, augmented reality interfaces, and image-editing workflows that require precise selection among similar items. Visionwhich challenges include robustly disambiguating in ambiguous scenes, handling noisy prompts, and maintaining efficiency for real-time use.

See also: visual question answering, referential expression comprehension, multimodal AI, interactive vision systems.