Not good for detecting small objects ? about asff HOT 10 CLOSED

@ruinmessi 1 more thing weird is that when I used the weights that I had trained, if I set the threshold score in range (0, 1) as usual, no predictions were shown. But when I set it 10000 smaller, the results appeared.

from asff.

Sorry, I have no idea about your dataset and training setting. As the img you showed is quite simple, I suggest you use a ''shallow'' model, like SSD, RFBNet for this task. They are more friendly for single gpu training, and the performance may be not too bad.

from asff.

@ruinmessi I know that my dataset is quite simple but the results I got from your model are not good. I had used Centernet before and I want to have a comparison between 2 models in some simple dataset first.

from asff.

@ruinmessi And tried to train the network again with coco 2014 valid, and it still detects very well. So I think this network is not good at detecting small objects as the fact that yolo itself is not good at handling that problem.
I also saw that in yolo_head you defined 9 anchors.
Do those anchors size effect the network when is trained for detecting small objects ?

from asff.

@trungpham2606 I need your test.py code for detecting the single picture,can you send it to me at my email [email protected] ,thank you

from asff.

@wyj18502986957 yeah, sure bro.

from asff.

@trungpham2606 I also need that test.py code, can you send it to me at my email [email protected]

from asff.

@trungpham2606 , @wyj18502986957 I want detecting my dataset,can you send the test.py code to me .my email is [email protected]. thanks

from asff.

@trungpham2606 What's the effect of centernet？

from asff.

@wushilian what do you mean "effect" ?

from asff.