Finding accurate location of objects from image API?

API
1

I am passing a JSON schema for output to the image API, asking for a list of objects with their description and bounding boxes. I get some bounding boxes that are somewhat related to the truth, but very loose. In Azure cognitive services I can get accurate bounding boxes of detected objects (but no custom prompt), so I am wondering if I could get that from here?

Example of what I get with bounding boxes from response JSON drawn onto image. You can see the bounding box has single decimals only. I use 512x512 images, so shouldn’t be scaling involved.

A modern kitchen features a gray palette with a sleek island and blue cushioned bar stools, highlighted by geometric colored frames. (Captioned by AI)
A modern kitchen features a gray palette with a sleek island and blue cushioned bar stools, highlighted by geometric colored frames. (Captioned by AI)736×666 103 KB

1 Like
2

For comparison, this is the results from Azure cognitive services

A modern kitchen features dark gray cabinetry, a herringbone tile backsplash, spherical pendant lights over a central island, and a blue stool with black legs. (Captioned by AI)
A modern kitchen features dark gray cabinetry, a herringbone tile backsplash, spherical pendant lights over a central island, and a blue stool with black legs. (Captioned by AI)1293×861 85.4 KB

3

Have you tried yolo?