Do you have object detection bounding box visualization results?

[c0nn3r]

Trying to debug my implementation's training on COCO, would love to see if you got this repo working on VOC.

[andreaazzini]

So, I'm still far from achieving good results on VOC (I had the time to complete only 72 epochs). Some examples (good, promising and bad).

My gut feeling is that something is still wrong in the encoder.

[andreaazzini]

@c0nn3r any better results on your side?

[c0nn3r]

Nothing yet, having trouble batching the bounding boxes. Does it only work when only one bounding box is needed to detect (or a lower number)? I'm thinking of writing tests for encoder.py and just brute force a working solution. I'm going to switch to getting VOC working.

[c0nn3r]

Also @andreaazzini, what are you using to produce the images shown?

[andreaazzini]

@c0nn3r https://github.com/andreaazzini/retinanet.pytorch/blob/328f63e12fdcd9e5b11f8828828b302245b13cf8/demo.py
I removed demo.py from master because I'm working on an evaluation script, and the demo script needs to be changed accordingly.

[c0nn3r]

@andreaazzini I took a look at the loss function and found some problems:

• I don't think we are normalizing focal loss over the number of background detections / ground-truth boxes (it was italicised in the paper, so it must be important!).
• The paper claims to have ~100K anchor predictions for the box subnet, that doesn't seem to match our ~47961.

[andreaazzini]

@c0nn3r Are you using my encoder or @kuangliu's?

[c0nn3r]

@andreaazzini I'm using your encoder