I sent an image to over a dozen LLMs that support vision, asking them:
Detect objects in this 1280×720 px image and return their color and bounding boxes in pixels. Respond as a JSON object: {[label]: [color, x1, y1, x2, y2], β¦}
None of the models did a good-enough job. It looks like we have some time to go before LLMs become good at bounding boxes.
I’ve given them a subjective rating on a 1-5 scale below.
Model | Positions | Sizes |
---|---|---|
gemini-1.5-flash-001 | π’π’π’π΄π΄ | π’π’π’π’π΄ |
gemini-1.5-flash-8b | π’π’π’π΄π΄ | π’π’π’π΄π΄ |
gemini-1.5-flash-002 | π’π’π΄π΄π΄ | π’π’π’π΄π΄ |
gemini-1.5-pro-002 | π’π’π’π΄π΄ | π’π’π’π’π΄ |
gpt-4o-mini | π’π΄π΄π΄π΄ | π’π’π΄π΄π΄ |
gpt-4o | π’π’π’π’π΄ | π’π’π’π’π΄ |
chatgpt-4o-latest | π’π’π’π’π΄ | π’π’π’π’π΄ |
claude-3-haiku-20240307 | π’π΄π΄π΄π΄ | π’π’π΄π΄π΄ |
claude-3=5-sonnet-20241022 | π’π’π’π΄π΄ | π’π’π’π΄π΄ |
llama-3.2-11b-vision-preview | π΄π΄π΄π΄π΄ | π΄π΄π΄π΄π΄ |
llama-3.2-90b-vision-preview | π’π’π’π΄π΄ | π’π’π’π΄π΄ |
qwen-2-vl-72b-instruct | π’π’π’π΄π΄ | π’π’π΄π΄π΄ |
pixtral-12b | π’π’π΄π΄π΄ | π’π’π’π΄π΄ |
I used an app I built for this.
Here is the original image along with the individual results.
Update
Adding gridlines with labeled axes helps the LLMs. (Thanks @Bijan Mishra.) Here are a few examples: