What could be reason for getting 0 as boxloss in SSD?

Gluon
1

I am trying to SSD network on my own dataset. However, I keep getting boxloss to be 0. Classification loss is not 0. But only box loss is 0.

When my default anchors, labels and class_predictions go through MutliboxTarget, I get 0 for box_target, box_mask, and class_target.

I have tried lot of things

  • Changed size of image
  • Converted image from rectangular to sqare
  • Tried with similar small dataset
  • Random translation and others.
  • Even passed some other dataset through my network (Where the network works fine and Multibox target returns some values apart from 0)
  • Tried classifying on same dataset with 2 classes. (Background, Object)
  • Let it run for long number of ephoces. (But still box loss is 0, class loss approaches 0 and eventually network gets converged only on class loss.
  • Ensured that training_target function works fine (copy pasted the literal code from MXNet object detection module)
  • Even tried it on small network with no body, still mutilbox target returns 0.]
  • Played around with lots of different anchorsizes and ratios.

So my question is what could be wrong? What could be wrong with my dataset? Why does multibox target always return 0 for everything?

And what exactly does a multibox target function do?

2

I meet the same problem two days ago. I found the problem is that the shape of label for every sample in one batch is different. Did you check that?

3

Yes my labels are padded to 117. Every label. So I have 117 label per image. (Batch size, 117, 5)

4

Are you using SSD from gluon CV or your own implementation?

5

My own implementation.

6

Is your “boxloss” an L1 loss between predicted box center and width/height and anchors? If so, did you try outputting these values for one sample to see what the output it?

7

My box loss is L1 loss.

However the issue is when I use MultiboxTarget (Before to computer targets for the purpose of calculating the boxloss) I get 0 matrix for box mask and target. This leads to overall box predictions to be 0.

Hence boxloss is 0.

Now what I don’t get is why am I getting MultiboxTarget’s output 0 no matter which image I pass.

The entire network minimizes the class loss and converges. What are the possible scenarios when MultiboxTarget could return 0 for box_mask and box_target?

Can you explain what the function exactly does then if there is a possibility I might can implement that.

8

Question to general audience are we suppose to normalize the labels between 0.0 and 1.0 to be able to use the MultiBoxTarget function?

Because when I normalized the labels between those values I was able to get non zero bounding box loss.