when inquiring about the relative positional realtionship of items, the returned results are always incorrect.

[zmf2022]

No description provided.

[RussRobin]

Hi @zmf2022 ,

Thank you for your interest in our work. You may want to provide the rgb, depth and text inputs so we can further assist you. Greatly appreciated!

[zmf2022]

[zmf2022]

the question: orange and apple, which one is closest to me?

[RussRobin]

Let's test if the model can recognitize the objects first. Can you try 3 QAs in a sequence?
-What is the bounding box of the object: orange?
-What is the bounding box of the object: apple?
-Which is closer to the camera, orange or apple?

[zmf2022]

ok, i will try, thanks!

[ThePassedWind]

Let's test if the model can recognitize the objects first. Can you try 3 QAs in a sequence? -What is the bounding box of the object: orange? -What is the bounding box of the object: apple? -Which is closer to the camera, orange or apple?

Can I input masks and semantic label instead of asking QAs like -What is the bounding box of the object?

[RussRobin]

You can tell the model the bounding boxes of objects you are interested in.

The reason for SpatialBot to fail your original question may be:

1. It does not recognize the objects as they are small. So adding bounding boxes or masks would help.
2. The confusion of 'me' with the people in your image. If you replace 'me' with 'camera', SpatialBot may be able to answer it correctly.

Hope it makes sense to you!