facebookresearch
Image as Exemplar: Tracks any object in an input image or video that has features similar to those of an Exemplar image.
I want to use a reference image as a prompt, similar to a text prompt, to track objects that provide the same semantic meaning as the reference image. Is this possible?
Hi,
I am encountering the same issue.
A section from the official blog on SAM 3 states: "SAM 3 supports a variety of prompt modalities, including both concept prompts such as simple noun phrases and image exemplars, as well as visual prompts, such as masks, boxes, and points [...]. This increases the flexibility and usability of segmentation, particularly for concepts that are rare or hard to describe with text alone." To me, this implies that we should be able to use a reference image (or "image exemplar") as a prompt, but I cannot find this functionality in the code. Am I missing something?
I have also ran into this issue. I can't get an exemplar bounding box to work with video propagation. Text only prompts work fine of course.
I agree. It appears that the current visual prompt implementation follows the SAM 2 logic (spatial/interactive refinement) rather than the semantic 'concept' logic described in the blog.
Has anyone successfully used a bounding box or image crop as a semantic exemplar to track a concept, rather than just as a spatial locator?
Has anyone figured out if SAM3 lets you do “few-shot” prompting by giving it reference images? I've just had success with text prompts
