YOLO version 3 equivalent in Keras

[hiba007]

Can anybody help me in building YOLOv3 architecture in keras? here is the link I think its residual network https://github.com/pjreddie/darknet/blob/master/cfg/yolov3.cfg @experiencor

[experiencor]

@hiba007 ResNet backend is currently supported. Please check the readme. If possible, can you provide a drawing of YOLOv3.

[ghost]

after reading, i found some main different things btw 2 and 3:

• new backbone network (similar to resnet)
• using feature pyramid as main concept (similar to retinanet, dssd, fpn)
• using anchors' size from input size, not output size (yolo2) -> new anchors 's size = 32x
• at each feature map of pyramid, use 3 different anchors (different scales) -> total 9 anchors on 3 levels of pyramid -> reduce computation, instead of 5 or 9 anchors at every level
• remove rescore iou, so label of positive bounding boxes is 1, not overlaps, change iou threshold to 0.5 because 0.5 is threshold for evaluation

[hiba007]

drawing means architecture ? here is cfg file https://github.com/pjreddie/darknet/blob/master/cfg/yolov3.cfg

[hiba007]

you mean this? @experiencor

[experiencor]

@hiba007 @txdat Interesting changes! I'll take a closer look.

[cthacker]

Yolov3 also doesn't use softmax and just uses logistic units for each class (multiclass classification)

[PythonImageDeveloper]

@txdat , what's this mean : using anchors' size from input size, not output size (yolo2) -> new anchors 's size = 32x i know about anchors and using k-means for getting these value but i don't know what's mean anchors from input and output ? that's mean in the v3 the anchors directly get from images ?

[experiencor]

in version 3, anchor sizes are absolute, not relative to grid size like in version 1 and 2

[PythonImageDeveloper]

what are the advantage of this method rather relative?

[experiencor]

@zeynali I'm not very clear about this. Still working on the new loss function.

[experiencor]

@hiba007 @txdat @zeynali YOLOv3 detection is out. Still working on the new multi-scale loss training.

[PythonImageDeveloper]

@experiencor , Thanks a lot , When will the multi-scale training be prepared? please explain about differences between v3 and v2 ? more details if possible. Thanks

[cthacker]

@experiencor New model looks good.

why do you resize the image but keep the aspect ratio the same and embed into a letterbox? Do you find that helps with accuracy?

Also are you planning to implement the lossfn for this network or no? I'm still trying to figure out everything that is different compared the standard yolo loss.

[PythonImageDeveloper]

@cthacker , Is not this implementation of yolo3 loss function corresponding to original Version 3?

[ghost]

@experiencor 2 things i dont understand in yolo3_detect.py: 1) netout[..., 5:] = netout[..., 4][..., np.newaxis] * netout[..., 5:]' why do you multiply iou_score to class_score? 2) you compute overlaps in (0. 1) scale not input's scale?

And prediction boxes are computed like yolo2, why dont warmup need in yolo3? prediction bbox have xy_pred in output's scale and wh_pred in input's scale?

[experiencor]

1. It's P(Class) = P(Class|Object) * P(Object)
2. When you compute IOU, it does not matter.
3. I don't why warmup is omitted in the new loss function. Asked a question but not yet get the answer.

[experiencor]

https://github.com/experiencor/basic-yolo-keras/tree/master/experimental => YOLO 3

[muye5]

Expect the training part for v3....deeply.

[experiencor]

It's here https://github.com/experiencor/keras-yolo3