Running Cube-RCNN on a different dataset

[johannes-tum]

Dear Authors,

I like your approach a lot. Thanks a lot also that you provide such an extensive documentation.
Especially that you define how to define the learning parameters in case of a different batch size.

Let's assume you have a different joint training dataset with about half the size. What would you change?
I would assume that you also linearly scale:

• SOLVER.STEPS
• SOLVER.WARMUP_ITERS
• SOLVER.MAX_ITER
• TEST.EVAL_PERIOD
• VIS_PERIOD
  Anything else?

Best wishes

Johannes

[mason1008]

Dear Johannes,

I have the same question as you. May I ask how you are currently approaching it?

Best wishes

Mason

[johannes-tum]

Hi Mason,
yes, I linearly scaled everything. However, In general, I find it quite challenging to get the OMNI3D code to run in stable way. That's the main challenge I encountered.
Best wishes
Johannes