START

This code is implemented base on the original code Yolov3 Tensorflow. The main purpose is improving the performance and can using on some Edge device

Clone this file

$ git clone https://github.com/buiduchanh/TF_yolov3.git

Train your own dataset

1. Prepare Dataset

Two files are required as follows:

• dataset.txt:

xxx/xxx.jpg 18.19,6.32,424.13,421.83,20 323.86,2.65,640.0,421.94,20 
xxx/xxx.jpg 48,240,195,371,11 8,12,352,498,14
# image_path x_min, y_min, x_max, y_max, class_id  x_min, y_min ,..., class_id 
# make sure that x_max < width and y_max < height

• class.names:

person
bicycle
car
...
toothbrush

2. Caculate Anchor

We provided file kmeans.py for caculate the anchor like coco_anchor.txt

3. Change config

Then edit your ./core/config.py to make some necessary configurations

__C.YOLO.CLASSES                = "./data/classes/voc.names"
__C.TRAIN.ANNOT_PATH            = "./data/dataset/voc_train.txt"
__C.TEST.ANNOT_PATH             = "./data/dataset/voc_test.txt"

If you want to use MobileNetV2 as backbone instead of Darknet53 just set the parameters in config same as below
__C.YOLO.BACKBONE_MOBILE        = True
__C.YOLO.GT_PER_GRID            = 3

Here are two kinds of training method:

4. Training

$ python train.py
$ tensorboard --logdir ./data

Result

We will update this result asap

Improve from original code

• MobileV2
• DarkNet
• Using Focal loss
• Added Batch Normalize

TODO

• Convert model to using in edge device
• Adding channel prunning
• Using Diou loss instead of Giou loss ( increase mAP ~5%)
• Adaptively spatial feature fusion ASFF which increase the mAP ~ 10%

Reference

Stronger-Yolo
focal-loss
kl-loss
YOLOv3目标检测有了TensorFlow实现，可用自己的数据来训练
Implementing YOLO v3 in Tensorflow (TF-Slim)
Understanding YOLO