Hello, everyone. I’m gonna train a network in PyTorch to detect small size object like cellphone.I have about 3k target images(Well, a small dataset). And after training, the precison is about 50%(sadly low, and I’ve done the data augmentation like image flip and so on). I have tried methods as much as I know, but it still doesn’t work out.Anyone who has such experience could give me some advice? Thank you so much!
Why not initialize your model using pretrained imagenet ?
I have tried transfer learning, but it worked poorly, cause there is so so litter cellphone pictures.
Which network are you using?
Can you classify properly good those items?
Can you classify properly good those items without contextual information?
Some detectors work better than others on small objects, e.g., RefineDet works better than YOLO.
You should also tune the anchor sizes accordingly and maybe use larger input image resolution.
I’m using yolo v3 and I have done the data preprocessing.
Ok, 3Q, I will have a try.