Bounding boxes smaller than expected with yolov4-tiny (3 yolo-layers)

[vinorth-v]

Hello,

I don't understand, the results are perfect with my darknet model but after converting to TF, the detection results are 2-3 times smaller for the bounding boxes.

I am using this config file https://github.com/AlexeyAB/darknet/blob/master/cfg/yolov4-tiny-3l.cfg

python detect.py --framework tflite --score 0.5 --weights /home/vinorth/Desktop/saved_models/yolo_models/3_yolo_layers/yolov4-tiny-3l_v2.tflite --imag
e /home/vinorth/Desktop/dataset/MINI_DATASET/images/folder_test/img_986.jpg --tiny --size 608

By running the command above, the bounding boxes are way smaller like these:

Do I need to change something in your code? How should I modify the anchors?

Thanks

[vinorth-v]

Should I change something here or somewhere else?

config.py

#YOLO options
__C.YOLO                      = edict()

__C.YOLO.CLASSES              = "./data/classes/obj_old.names"
__C.YOLO.ANCHORS              = [12,16, 19,36, 40,28, 36,75, 76,55, 72,146, 142,110, 192,243, 459,401]
__C.YOLO.ANCHORS_V3           = [10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326]
__C.YOLO.ANCHORS_TINY         = [23,27, 37,58, 81,82, 81,82, 135,169, 344,319] # yolov4-tiny-2l.cfg (mask = 1,2,3)
#__C.YOLO.ANCHORS_TINY         = [10,14, 23,27, 37,58, 81,82, 135,169, 344,319] # yolov4-tiny-2l-custom.cfg (mask = 0,1,2)
#__C.YOLO.ANCHORS_TINY         = [12,16, 19,36, 40,28, 36,75, 76,55, 72,146, 142,110, 192,243, 459,401] # yolov4-tiny-3l-custom.cfg (mask = 0,1,2) -> not working
__C.YOLO.STRIDES              = [8, 16, 32]
__C.YOLO.STRIDES_TINY         = [16, 32]
__C.YOLO.XYSCALE              = [1.2, 1.1, 1.05]
__C.YOLO.XYSCALE_TINY         = [1.05, 1.05] #yolov4-tiny-2l
#__C.YOLO.XYSCALE_TINY         = [1.05, 1.05, 1.05] #yolov4-tiny-3l-custom.cfg -> not working
__C.YOLO.ANCHOR_PER_SCALE     = 3
__C.YOLO.IOU_LOSS_THRESH      = 0.5

yolov4.py

def YOLOv4_tiny(input_layer, NUM_CLASS):
    route_1, conv = backbone.cspdarknet53_tiny(input_layer)

    conv = common.convolutional(conv, (1, 1, 512, 256))

    conv_lobj_branch = common.convolutional(conv, (3, 3, 256, 512))
    conv_lbbox = common.convolutional(conv_lobj_branch, (1, 1, 512, 3 * (NUM_CLASS + 5)), activate=False, bn=False)

    conv = common.convolutional(conv, (1, 1, 256, 128))
    conv = common.upsample(conv)
    conv = tf.concat([conv, route_1], axis=-1)

    conv_mobj_branch = common.convolutional(conv, (3, 3, 128, 256))
    conv_mbbox = common.convolutional(conv_mobj_branch, (1, 1, 256, 3 * (NUM_CLASS + 5)), activate=False, bn=False)

    return [conv_mbbox, conv_lbbox]

[aditya-5842]

I am facing the same issue for yolov4-tiny-3l model. How did you solve this problem?

[vinorth-v]

Hi @aditya-5842, unfortunately not, still facing this issue. And it seems this repo is not going to be updated anymore... @hunglc007

[aditya-5842]

@vinorth05 After working 6-8 hrs, I am able to convert yolov4-tiny-3l from darknet framework to `TensorFlow framework. Now it is working for me.

There are many things are required to make it work. Soon I will update this repo by forking it.

[aditya-5842]

@vinorth05 You can check this branch. It is working for me. https://github.com/aditya-5842/tensorflow-yolov4-tflite/tree/yolov4-tiny-3l-fix

You can see the necessary changes here.

[vinorth-v]

Hi @aditya-5842, thank you for sharing this! I will try your changes when I have time and get back to you!

Edit: It works great, thank you!

[vinorth-v]

However, it is strange that when I compare results with my test dataset, yolov4-2l model has better results than yolov4-3l model. (It is supposed to be the opposite, like when I tested with darknet)

Do you think you missed some parameters?

[aditya-5842]

Yes, YOLOV4-Tiny-3l is supposed to give better results than YOLOV4-Tiny (2l). I trained both models for detecting a single class and I got better results with YOLOV4-Tiny (2l).

I think I have handled the parameters carefully. In my case result of darknet and TensorFlow were same (for images).

Can you share the examples where TensorFlow output is different than darknet output.