YOLOP
YOLOP copied to clipboard
hustvl
Anchor setup
Dear Authors,
thanks for your great work. I have a basic question: https://github.com/hustvl/YOLOP/blob/a30c32c3ea17970e8daa94ee99e5803acdcf1a48/export_onnx.py#L44 I believe this is your default anchors, [3, 9, 5, 11, 4, 20], [7, 18, 6, 39, 12, 31], [19, 50, 38, 81, 68, 157] My question is why the anchors are vertical instead of horizontal (for horizontal vehicles) ? stride 8: [3, 9, 5, 11, 4, 20]: stride 16: [7, 18, 6, 39, 12, 31], stride 32: [19, 50, 38, 81, 68, 157] I think the order is [anchor1_w, anchor1_h, anchor2_w, anchor2_h, anchor3_w, anchor3_h] for each layer https://github.com/hustvl/YOLOP/blob/a30c32c3ea17970e8daa94ee99e5803acdcf1a48/lib/models/common2.py#L241
One reason I could think of is maybe the autoanchor makes 1280x720 -> 640x640 (not keeping aspect ratio), and thus make the anchors more vertical-like (but I have not yet look into the code) But the inference in YOLOP seem to be keeping aspect ratio.
Above is my question. Thanks again for the great work.
I could think of another reason is that maybe person class is used in autoanchor instead of vehicle class: I found yolov5 is with some horizontal anchors, too: https://github.com/ultralytics/yolov5/blob/master/models/yolov5s.yaml#L7
- [10,13, 16,30, 33,23] # P3/8
- [30,61, 62,45, 59,119] # P4/16
- [116,90, 156,198, 373,326] # P5/32
Adding comment: the filtered shall be done in bdd.py already: https://github.com/hustvl/YOLOP/blob/bff896e34d1c6ed75ea40353184e2ea8058b8a93/lib/dataset/bdd.py#L40, so only car, truck, bus, train are remaind in the bbox db.
Dear Author:
I just found the bug, if you print out:
https://github.com/hustvl/YOLOP/blob/main/lib/utils/autoanchor.py#L87 " shapes = img_size * dataset.shapes / dataset.shapes.max()"
you will find the shapes is [360, 640], but you multiple it with w and h: in next line wh0 = labels[:, 3:5] * shapes this is to say, you shall make the shapes from [360,640] to [640,360].
Please confirm whether I am correct. Thank you
