YOLOv3-tiny-custom-object-detection

As I continued exploring YOLO object detection, I found that for starters to train their own custom object detection project, it is ideal to use a YOLOv3-tiny architecture since the network is relative shallow and suitable for small/middle size datasets

Environment:

OS: Ubuntu 16.04
CUDA 9.0
cuDNN 7.0
Tensorflow 1.8.0
OpenCV 3.3.0

GPU:Nvidia Geforce GTX 1080

Great sources:

https://github.com/AlexeyAB/darknet

https://github.com/AlexeyAB/Yolo_mark

Step 1:

Go to YOLO website https://pjreddie.com/darknet/yolo/, follow the instructions and have your Darknet installed.

Step 2:

Compiling with CUDA and OpenCV, here is the instruction: https://pjreddie.com/darknet/install/#cuda Make sure your can run those commands before training your own dataset

./darknet detector demo cfg/coco.data cfg/yolov3.cfg yolov3.weights

./darknet detector demo cfg/coco.data cfg/yolov3.cfg yolov3.weights

Step 3:

Prepare your own labled dataset: Here is a great tool I used to annotate my own datasets: https://github.com/AlexeyAB/Yolo_mark Remember, dataset is very crucial for deep learning and your model's performance, it is time consuming to prepare your own annotated dataset.

Step 4:

Copy all your prepared images and their corresponding .txt files to the directroy in your Darknet file:

./home/Darknet/data/img

Step 5:

In Step 3, you should have created a txt file called : obj.names. Since I trained on 6 objects, I named my file as obj6.names and they are: babybuggy, bicycle, fire_engine, motor_cycle, scooter, moter_scooter. I have uploaded a sample obj.names file in this post.

Step 6:

Create a "train.txt" file which including the directory and the names of your labeled images for training:

data/img/n02834778_12542.jpg
data/img/n02834778_12545.jpg
data/img/n02834778_12553.jpg
data/img/n02834778_1255.jpg
data/img/n02834778_12592.jpg
data/img/n02834778_12604.jpg
data/img/n02834778_12605.jpg
data/img/n02834778_12643.jpg
data/img/n02834778_12654.jpg
data/img/n02834778_12655.jpg
data/img/n02834778_12658.jpg
data/img/n02834778_12673.jpg

Step 7:

Create a "test.txt" file which including the directory and the names of your labeled images for testing, remember do not test on the same images which used for training already.

data/img/n02834778_1313.jpg
data/img/n02834778_1325.jpg
data/img/n02834778_1339.jpg
data/img/n02834778_1362.jpg
data/img/n02834778_1365.jpg
data/img/n02834778_1372.jpg
data/img/n02834778_1380.jpg

Step 8:

In Darknet file, there is a cfg file. Creat your own YOLOv3-tiny.cfg in that file. I have posted a sample "YOLOv3-tiny6.cfg" file which I used to train my model.

Step 9:

You can download a pretrained weight from: http://pjreddie.com/media/files/darknet53.conv.74

Step 10:

Now you can start your training by typing: ./darknet detector train data/obj6.data yolov3-tiny6.cfg darknet53.conv.74
The trained weights will be saved at your ../darknet/backup

Step 10:

After you monitored the training for maybe 10,000 iterations, you can stop training and test out your model by typing:
./darknet detector test data/obj6.data yolov3-tiny6.cfg backup/yolov3-tiny6_10000.weights data/test/babybuggy.jpg