Light architectures for object detection

[valery-kustov]

Hi, Markus,

Cool repo for detection models.
I'm interested in more light architectures for object detection and tried to change VGG16 (SSD512_body and SSD300_body) to mobile_net or mobile_net_v2.
But I can't get an adequate quality for them on VOC dataset.
Metrics on 30 epoch:
loss 4.34212 conf_loss 0.02317 loc_loss 2.02536
precision 0.54101 recall 0.17049 fmeasure 0.25575 accuracy 0.86921
val_loss 4.28855 val_conf_loss 0.02258 val_loc_loss 2.03045
val_precision 0.49608 val_recall 0.16673 val_fmeasure 0.24578 val_accuracy 0.87617

Probably it's due to new feature maps or incompatibility between backbone and extra blocks or something in training pipeline.
Did you try to do the same? If yes, could you give your contact to discuss how to use these architectures in your repo?

Best Wishes,
Valery

[mvoelk]

I haven't spent that much time training SSD. As far as I know, the VGG16 models are pre-trained on ImageNet. you could try the backbone in ssd_model_separable.py, with more features and more training data (MS COCO for instance.). I'm also not sure which parameters for data augmentation are best.

[mvoelk]

See also #21 #37 I posted some lightweight architectures there.

[valery-kustov]

Thank you. I'll train some models within few days and inform you about results.

[valery-kustov]

Hi,

When you change backbone neural network to mobile net v1 (as an example) PriorUtil.encode() starts to exclude some initial samples (line 548 PriorUtil.encode(y) from ssd_data.Input_Generator). Therefore the network could not be trained correctly.
The reason is in iou function, overlap threshold and priors_norm, even with threshold==0.01 we lose some samples.

But I don't know the full logic of PriorMap and PriorUtil to understand, what's wrong.
I tested your realisation of the network from #21 (comment)
the error is kept.

For SSD300, SSD512, ssd_model_separable PriorUtil.encode() works well.

Could you support?

[mvoelk]

Okay, this is known issue to me. I'm not sure what's wrong with the original assignment (ground truth to prior boxes). I investigated different assignment methods, which led to the condition in line 405.

You can set model.ssd_assignment = False or pass ssd_assignment=False to the PriorUtil.

I used the plot_assignment method to debug this behavior.