YoloV3ForCustomDataset

Install ffmpeg using : conda install -c conda-forge ffmpeg
Go to the directory of the video and type ffmpeg -i filename.mp4 -vf fps=4 image-%d.jpeg
Extracted images will be the in directory of the video
Draw bounding boxes around the extracted video using LabelImg
Prepare the dataset in YOLO Format using 03DataPreparation Notebook

Section 4 : Installing OIDv4 toolkit for downloading images from a huge dataset

Clone the OIDv4 Repository OIDv4
Activate your python environment and navigate to the directory where the repository was cloned and type the following command : pip install -r requirements.txt
To verify type the following command : python main.py -h
After verifying type the following command : python main.py downloader --classes Car Bicycle_wheel Bus --type_csv train --multiclasses 1 --limit 9
9 Images of the classes Car Bicycle_wheel and Bus will be downloaded and two csv file will also be downloaded in the OID Folder.
One CSV file will contain annotations and the other will contain and the class description.
In order to verify type the following command : python main.py visualizer
Rename the Car_Bicycle wheel_Bus folder to Car_Bicycle_wheel_Bus to avoid errors
Convert annotations from the csv file using 04ConvertingAnnotations notebook
For easier access I would have copied the Dataset and csv_folder Folders to Section 4 and would have deleted the OIDv4 folder
Prepare data in YOLO Format using 04DataPreparation notebook

Section 4b : Joining the labelled and custom datasets

Copy only the images that we downloaded from the huge dataset Section4 and the images we extracted from the video Section 3 to a new folder LabeledCustom.
Run the 04JoiningDatasets in Section 4b to join the dataset and prepare it in YOLO Format.

Section 5 : Traffic Dataset

Download the traffic dataset from the link provided
Convert annotations from the csv file using 05ConvertingAnnotations notebook
Prepare data in YOLO Format using 05DataPreparation notebook

Note : In all notebooks use the proper directory

Note : In Sections 3,4,4b,5 if you want to check if we have prepared the images correctly create a classes.txt file in the same order and open LabelImg and verify

Section 6 : Darknet

Install Tensorflow with GPU Support using this tutorial from Jeff Heaton Tensorflow GPU
Install darknet using the this tutorial Darknet Installation
Create the folder weights and add yolov3.weights
In the data folder in the darknet directory add the test-image.jpg and test-video.mp4 files

Command to check installation of darknet

For image file

Go to darknet root directory and type : darknet.exe detector test cfg\coco.data cfg\yolov3.cfg weights\yolov3.weights -thresh 0.85 -ext_output data\test-image.jpg

Output will be stored as predictions.jpg

For video file

Go to darknet root directory and type : darknet.exe detector demo cfg\coco.data cfg\yolov3.cfg weights\yolov3.weights -thresh 0.85 -dont_show data\test-video.mp4 -out_filename result.avi

Output will be stored as result.avi

For real time camera

Go to darknet root directory and type : darknet.exe detector demo cfg/coco.data cfg/yolov3.cfg weights/yolov3.weights -thresh 0.85 -c 0

After checking if darknet works

Copy the ts_data.data from the Traffic Dataset and custom_data.data from Car_Bicycle_wheel_Bus dataset to the cfg folder in the darknet directory
Open the yolov3.cfg file in the cfg folder in the darknet directory and copy its contents to four different files namely yolov3_ts_train.cfg, yolov3_ts_test.cfg, yolov3_custom_train.cfg, yolov3_custom_test.cfg and save it in the cfg folder.
Open the two files for training, uncomment lines # batch=64 # subdivisions=16 and change the value of batch to 64 and the value of subdivisions to 64. Delete the lines batch=1 , subdivisions=1 . Save results.
Open two files for testing and delete lines # batch=64 , # subdivisions=16 . Save results.
Next is changing max_batches (i.e) total number of iterations for training and the general formula used is max_batches = classes * 2000 (but not less than 4000)
Steps are calculated as 80% and 90% from max_batches.
For example, if number of classes is equal to 2, then:
- max_batches=4000
- steps=3200,3600
Next, it is needed to update number of classes in every of three yolo layers in the end of the configuration files. Also, it is needed to update number of filters in convolutional layers right before such every yolo layers but not anywhere else. It is needed in order to properly connect convolutional layer that is right before yolo layer in accordance with number of classes in dataset.
General equation that represents how to calculate proper number of filters in three convolutonal layers right before every of three yolo layers is as following
For example, if number of classes is equal to 2, then:
- filters = (classes + coordinates + 1) x masks
- For example, for COCO dataset it will be as following: filters = (80 + 5) x 3 = 255

Training

For Traffic Dataset type the following command : darknet.exe detector train cfg\ts_data.data cfg\yolov3_ts_train.cfg weights\darknet53.conv.74 -dont_show
For Car Bicycle_wheel Bus Dataset type the following command : darknet.exe detector train cfg\custom_data.data cfg\yolov3_custom_train.cfg weights\darknet53.conv.74 -dont_show

Note : In all Data Preparation Notebooks give the proper location of the backup location

Weights will be saved every 100 iterations and 1000 iterations. To define the best one to use for detection to make sure no overfitting occurs we use the mAP algorithm
The goal is to find weights that have the biggest mAP.
For example, if we consider Traffic Signs dataset, and if mAP for 7000 iterations is bigger than for 8000, then it is needed to check weights for 6000 iterations. Next, if mAP for 6000 iterations is already less than for 7000 iterations, then you can stop checking and use weights for 7000 iterations in detection tasks.
Also, it is possible to continue checking weights between 6000 and 7000 iterations, trying to find weights even with bigger mAP.
To find the mAP for weights of 8000 iteration for the Traffic Dataset go to the darknet directory and type : darknet.exe detector map cfg\ts_data.data cfg\yolov3_ts_train.cfg backup\yolo-obj_8000.weights.
To find the mAP for weights of 6000 iteration for the Traffic Dataset go to the darknet directory and type : darknet.exe detector map cfg\custom_data.data cfg\yolov3_custom_train.cfg backup\yolo-obj_6000.weights.

Testing

Copy buses-to-test.jpeg and buses-to-test.mp4 file to the data folder in the darknet directory.
Copy traffic-sign-to-test.jpeg and traffic-sign-to-testmp4 to the data folder in the darknet directory.
To test the Traffic Dataset Image type the following command : darknet.exe detector test cfg\ts_data.data cfg\yolov3_ts_test.cfg weights\yolov3_ts.weights -ext_output data\traffic-sign-to-test.jpg
Output will be stored as predictions.jpg
To test the Car Bicycle wheel Bus Dataset Image type the following command : darknet.exe detector test cfg\custom_data.data cfg\yolov3_custom_test.cfg weights\yolov3_custom.weights -ext_output data\buses-to-test.jpg
Output will be stored as predictions.jpg
To test the Traffic Dataset mp4 type the following command : darknet.exe detector test cfg\ts_data.data cfg\yolov3_ts_test.cfg weights\yolov3_ts.weights -thresh 0.85 -dont_show data\traffic-sign-totest.mp4 -out_filename result.avi
Output will be stored as result.avi
To test the Car Bicycle wheel Bus Dataset mp4 type the following command : darknet.exe detector test cfg\custom_data.data cfg\yolov3_custom_test.cfg weights\yolov3_custom.weights -thresh 0.85 -dont_show data\buses-to-test.mp4 -out_filename result.avi
Output will be stored as result.avi

To improve model accuracy

Set the random parameter to 1 : 3 times in every yolo layer inside cfg file for training.
height=608 , width=608 : Increase resolution for input images inside cfg files for training and testing. Choose any number that is divisible by 32 without leaving a remainder.
max_batches : Double or triple total number of iterations inside cfg file for training. Don't forget also to update parameter steps giving 80% and 90% accordingly.
Increase total number of labelled images in dataset. As well as, add empty images without any needed objects.

MBadriNarayanan/YoloV3ForCustomDataset

YoloV3ForCustomDataset

Done in Windows Machine

Resources not included in the repository due to size

Important commands for dataset preparation and execution and checking the result

To Prepare Images in YOLO Format

Section 3 : Extracting Frames from Video