how can I increase small objects'mAP?

[joshiaLee]

Hi
I selected 10 small objects in SUNRGBD and I made those trained in this model.
but, result showed only 0.1mAP( learning_rate = 0.001, epoch = 12, Objects were cup,bottle,bowl,book etc)
I guess this model was designed for big objects not small things, right?
then, how can I increase small objects' mAP? are there any parameters to do that?

[filaPro]

Hi @joshiaLee ,

First, if I remember correctly the main problem here is that there are extremely low examples of cup, bottle, bowl, book, ... in the train and validation splits. So, the quality of the detection will not be as good as for example chair category.

The second thing is the voxel_size parameter in configs. With the default value of 0.01 the objects are assigned on 4 feature levels with steps 8cm, 16cm, 32cm and 64cm. For small object you may need smaller voxel size, i.e. 0.005 and less feature levels i.e. 2 like here.

[joshiaLee]

Thank you sooooo much!

[joshiaLee]

I'm compused a little
feature levels you meaned is n_scores=2? or n_reg_outs=6?
what both mean?

[filaPro]

Here we decrease n_outs, in_channels and n_scales.

[joshiaLee]

I got it
Thank you!