Small object detection

[yusiyoh]

I have images of size 2048*1024, and have labels with a width of 11 px on average. (DriveU Traffic Light Dataset). I want to train a model that is capable of detecting these traffic lights which can be categorized as small objects. Are there any structural changes for yolov4-csp as in yolov4 version of AlexeyAB (changing some layers and strides) ?

[WongKinYiu]

yes.

[yusiyoh]

So, what should I change in cfg?

[amkonshin]

I have the same task, but my objects a bit bigger. I assume u need to change anchor boxes. @WongKinYiu do we need to change the architecture of the net to fit our task well?

[yusiyoh]

I have the same task, but my objects a bit bigger. I assume u need to change anchor boxes. @WongKinYiu do we need to change the architecture of the net to fit our task well?

I think the model automatically generates appropriate anchors before training for custom datasets. However, still for better performance, there are some changes in the darknet repository (in cfg file).

[engrjav]

@yusiyoh and @WongKinYiu hey i also need to train my dataset on small objects. Can you tell if i have to modify the filters as suggested in darknet cfg file for YOLO v4. Right now my scaled YOLOv4p7 gives results much lesser in mAP than yolov5 and even CenterNet(objects as points)

[engrjav]

also i have input resolution of 416x416 but it scales up to 512 automatically which i think should not compromise performance. but since it does not resize my validation images can it still be an issue?