Hi, I trained on a new coco-like dataset with 7 classes but detection were all wrong

[lucasjinreal]

I try to detection trafficlights and trafficsigns, I only using coco config and it's labeling format just like coco, but after training the detection were all wrong.

the location loss around 3.8x, any insights on what's going wrong?

(training on coco actually works and get a very impressive results)

[PingoLH]

Hello Jinfagang, There could be many reasons for this. The loc loss is pretty large, so I guess the problem is with the bounding box regression. It might be helpful to check some places:

1. In layers/modules/multibox_loss.py, check that the size of pos after pos = conf_t > 0 is reasonable. It is the number of positive instances and shouldn't be zero or too large.
2. If the input image size is not 512 or 300, then you may need to modify the config file
3. In preproc (data/data_augment.py ), There is a normalization process for the boxes, make sure that you get the right shapes here. They shouldn't be too small (<0.01)

[lucasjinreal]

@PingoLH Thanks so much for advise. I will have an investigate .

[tuongtranngoc]

Hi @jinfagang, Do you have solution for training object detection? I also have same problem?