Object Counting

Object counting studies how to identify and quantify instances of a requested kind, including small, crowded, or partly occluded objects. Methods range from density-map regression and point localization to detection-and-tracking pipelines, while newer exemplar- and text-guided models use visual or language prompts to specify what to count. Vision-language models are also tested for counting and spatial reasoning, with benchmarks increasingly checking semantic grounding, distractors, counting granularity, and performance under occlusion or scene changes. A central challenge is making counts both accurate and faithful to the intended objects: models can recognize scenes yet confuse similar categories, merge instances, or fail to translate visual evidence into a number.