search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.67k stars 7.96k forks source link

Shallow model is sensitive to imbalance class samples #3883

Open erikguo opened 5 years ago

erikguo commented 5 years ago

I trained two models for the same data.

I found that shallow model is very sensitive to imbalance compared to deep model.

model: yolo_v3_tiny_pan3_aa_ae_mixup_scale_giou.cfg (mixup = 0), very sensitive

model: yolov3-spp-2-pan-scale.cfg, good results with even much imbalanced samples.

Do you have the same expierence?

AlexeyAB commented 5 years ago

model: yolo_v3_tiny_pan3_aa_ae_mixup_scale_giou.cfg (mixup = 0), very sensitive

model: yolov3-spp-2-pan-scale.cfg, good results with even much imbalanced samples.

What do you mean? Can you show APs of both models?

erikguo commented 5 years ago

I will try them again to do the comparison in the following week, and show you the results.

In my dataset, there are several classes sample is just one to ten of the others. But when I use tiny model training, the prediction is poor for these classes. But when using spp-pan-scale model traiing, the prediction is good.