YOLOv3 tensorflow

Build a real-time bounding-box object detection system for the boat (using fine-tuning in tensorflow based on YOLOv3-416 weights trained en COCO dataset). Then use my own data set for distinguish different type of boat

Update: This code was written during my internship in 2018 and is no longer maintained now!!!

20/05/2022: losing all stars after making this repo private

Inspired by YAD2K, Darknet

The full details are in this paper

Input to CCNs(Features block)	General	3 Scales	Features

Test

1. Clone this folder

2. Transfomer the pre-trained weights in Darknet to keras (may be skip this etape to etape 3)

   • wget https://pjreddie.com/media/files/yolov3.weights
   • python3 convert.py yolov3.cfg yolov3.weights yolov3.h5
   • python3 yolo.py

3. Or download the pre-trained weights in keras from here

4. Run python3 propagation.py

Results (La Rochelle, la belle ville :) )

YOLOv3-608	YOLOv3-416	YOLOv3-320

Train for your own dataset

1. Run python3 boat_annotation.py to get 3 files: bateau_train.txt, bateau_valid.txt, bateau_test.txt

   • In each file contains path_to_image obj1 obj2 ...
   • With obj1: x1_min, y1_min, x1_max, y1_max

2. Run python3 train.py

3. In propagation.py, modify classes_path to boat_classes.txt

4. Run python3 propagation.py

5. Enjoy your results!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!