How to solve the problem of shrinking?

[theodoruszq]

trafficstars

Hey, recently I am trying to reproduce your work, but after I train my model on Imagenet and ALOV. My model tends to shrink to a point or expand outside the image quickly, have you ever encounter this problem? How to solve it? I check every possible points in my code and visualize them but have no idea.

Thank you!

[subhadip3]

Hi, I am having the same problem. Did u solve it?

[theodoruszq]

Yes, I have already solved it.

[subhadip3]

Could you please let me know where the problem was and how you solved it?

[theodoruszq]

I am not sure whether my problem is same with you. But I think it could be such reasons following:

• Have you checked your box_scale_factor? It is important, and by Held's implementation, it is 10.
• Have you visualized your data layer, where always results in some weird problems?

If you have already checked the above problems carefully but the problem still exists, I think you could do complete unit tests for each module.