Anyway to get projected 2d bounding box coordinates from 3D space?

So I'm currently workign on a POC to see if we can use Unity to generate synthetic labeled data for our DL model to then train on. I've got an overall 3D environment with products on a shelf, but now I need to create labeled data output. I rotate the products and the cam around the shelf and take snapshots. I now want to label the products and find where the bounding boxes are for those products ("x", "y", "w", "h") for each image taken and store it in a JSON file.

I figured I would first see if anyone has seen anything already created out there before I go trying to figure out how to do this myself. If anyone has any info/advice, I would love to hear it. Thanks!

 

You can get the axis-aligned bounding box (AABB) from a collider or a renderer attached to the object. I'm guessing you have a MeshRenderer, so you can access meshRenderer.Bounds to get your bounding box.

I recommend you make sure you understand how an AABB works before you decide this will work for you because the box can resize when your object rotates. If this does not work for you, then you will probably want to loop through the verts in your object and find the Vector3 min and max in local space, then convert them to world space.

Once you've got your AABB or Vector3 min/max, use the camera that can see your 3D object to call camera.WorldToScreenPoint() to convert those 3D Vectors into 2D screen coordinates.

 

Thanks so much Gambit-MSplitz! This was the perfect solution and gave me what I was looking for.

 

I have similar issue, but i didn't understand the recommendation.
I am moving camera to various positions and capturing screenshot, I want to know the bounding box of textObject in the scene for each and every camera position. The scale of the text changes each time, the bounds value for MeshRenderer, I get from Unity and converts to WorldToScreenPoint() is not giving useful information. The value of bounds is more than actual image size.

how to solve this issue for scaling problem?

 

That is because it is an AABB

AABB's cannot be rotated, so when you rotate your image, the AABB actually becomes larger so that it can fit your whole rotated object inside without rotating itself. Here is an example image: the red box is the AABB, and even though the text is the same size in both places, the AABB is larger for the rotated text.



I'm afraid I don't know enough about what your needs are for the bounds that you're after. I recommend posting a new forum thread and briefly describing what you plan to do with the bounds or points you are trying to get.