About the anchors about pytorch-yolov3 HOT 3 CLOSED

I understand now. The network actually downsamples the image until the first prediction and then upsamples it until the last. So the first prediction is on the smaller feature map and the last one on the biggest feature map which explains why small objects are detected at the end (more information) and big objects in the begining. I know you mentioned it on your post @alexandru-dinu but I didn't notice at the time I read it.

from pytorch-yolov3.

1. For the first yolo layer, at the most coarse scale (13x13 for 416x416 image), the last 3 anchors are used:

mask = 6,7,8
anchors = 10,13,  16,30,  33,23,  30,61,  62,45,  59,119,  116,90,  156,198,  373,326

This makes sense, you want this layer to detect large objects in the image. The last yolo layer, at the finest scale (52x52 for 416x416 image), uses the first 3 anchors (smallest); this will detect smaller objects in the image.

2. If I recall correctly, anchors are bounding-boxes with (w, h) in [416, 416]. So they get normalized by dividing their sizes by 416.

from pytorch-yolov3.

But the receptive field of the first yolo layer is smaller compared to the one of the last layer so it would make sense to me that the first layer would detect the small objects. And because of the stride doesn't it make more sense that the first layer would detect small objects that the last one might miss since there is 32 pixel gap between the last yolo layer and the input image ?

from pytorch-yolov3.