1. Environment
2. [How2train](#training your own dataset)
3. what is yolov4
4. Evalute
5. Do in future

windows10(you also can run these codes on linux) tensorflow-gpu==1.13.1
keras==2.2.4

1、label you dataset in VOC format and use VOC development toolkit
2、put your image dataset in yolo3-keras-techi/VOCdevkit/VOC2088/JPEGImages
3、put your label files in yolo3-keras-techi/VOCdevkit/VOC2088/Annotations
4、run voc_annotation.py to generate true label files

classes = ["warning sign", "no reflective cloth", "reflective cloth", "staircase", "insulating tool", "tool"]

6、modify voc_classes.txt to class names in your project
7、run train.py

python train.py

8、tune parameters in configs.py to make your model be start-of-art

1、run get_gt_txt.py to get ground truth of valid dataset

python get_gt_txt.py

2、run get_dr_tex.py to get detecting results of valid dataset

python get_dr_tex.py

3、run get_map.py to get mAP0.5 of your model

python get_map.py

YOLOv3 is a object model as accuracy as SSD but 3 times faster. 1.YOLOv3 use independent logistic classifiers because a softmax is unnecessary for good performance.

2.YOLOv3 predicts boxes at 3 different scales. And use K-means clustering to determine our bounding boxes priors.

On the COCO datasets the 9 clustera were: (10× 13); (16× 30); (33× 23); (30× 61); (62× 45); (59×119); (116 × 90); (156 × 198); (373 × 326).

3.New feature extractor named DarkNet53.

1.We tried using the normal anchor box prediction mechanism where you predict the x; y offset as a multiple of the box width or height using a linear activation. We found this formulation decreased model stability and didn’t work very well. 2.Linear x; y predictions instead of logistic. We tried using a linear activation to directly predict the x; y offset instead of the logistic activation. This led to a couple point dop in mAP. 3.Focal loss:using focal loss will drop accuracy and i use it in my project where i find using focal loss does not help make my model better. 4.Dual IOU thresholds and truth assignment. Faster RCNN uses two IOU thresholds during training. If a prediction overlaps the ground truth by .7 it is as a positive example, by [:3−:7] it is ignored, less than .3 for all ground truth bjects it is a negative example. We tried a similar strategy ut couldn’t get good results.

• Transfer Learning
• multi-scale prediction
• Data Aug
• Diou
• Ciou
• Focal loss
• label smooth(i do not recommend to use label smooth in object detection, you can use it in single classifier task)
• ATSS Adaptive Training Sample Selection
• Mish
• MixUp
• Genetic Algorithm
• WBF(Weighted-Boxes-Fusion) postprocess https://arxiv.org/abs/1910.13302
• soft NMS
• Gaussian YOLOv3
• ASFF