YOLOV4-Tiny improved

---

Improved method 1
Activation function (LeakyReLU to Mish) improved

Improved method 2
Bayes modifier (include sigmoid and soft max) before NMS

Final Performance

How to run

(a) Inference

1. Download the weight in link https://drive.google.com/file/d/1YPMjwnqV4NJu3Lm2sTNpC3XwrQPJjMzx/view?usp=sharing, put the weight file into model_data.
2. run get_prob.py to generate bayes weight.
3. Then you can run the predict.py.
   The program may ask you the path of image, you can input img/street.jpg as you want.

(b) Test on Pascal VOC 2007+2012

1. Download the dataset with link https://drive.google.com/file/d/1OIRpaoKEGxrTUJ5JyYiGHKI19y1j2UHm/view?usp=sharing
2. Unzip the file in yolov4-tiny-improved
3. Change model_path and classes_path in yolo.py
   model_path is the path of weight file like model_data/final.pth classes_path should be the class nam.
4. run get_map.py

(c) Train on Pascal VOC 2007+2012

1. Data process run voc_annotation.py to generate index files.

2. Train run training.py

3. Test change the model_path in yolo.py (the path you trained can be find in log directory)
   run predict.py or get_map.py

Result show

Reference

https://github.com/bubbliiiing/yolov4-tiny-pytorch
https://github.com/AlexeyAB/darknet