Vehicle-Detection-Classification-YOLO-MobileNet

• Clone this repo

• Download yolo weights from https://pjreddie.com/darknet/yolo/ (Download TinyYOLO weights if you want to use that or else YOLO weights) For my work i have used TinyYolo as its faster and light weight.

• If you want to use TinyYOLO use command A else for YOLO command B

  A. python convert.py yolov3-tiny.cfg yolov3-tiny.weights model_data/yolo_tiny.h5

  B. python convert.py yolov3.cfg yolov3.weights model_data/yolo.h5

• To test: python yolo_video.py --image Then provide path to any test image

• Create a seperate environment to avoid any dependency clash conda env create -f test\dependecies.yml car_env conda activate car-project-env

• pip inststall --upgrade Pillow

• python MobileNet_TransferLearning.py (Feel free to change Optimizer/Epoch or any other ML technique according to your requirement) By Default I haev used Adam optimer with lr=0.0001 for 20 Epoch

• python VideoReader.py

Note: To know more about system please go through Vehicle_Detection_YOLO.pdf

Working Demo:

Base Structure of Mobilenet:

MobileNet Transfer learning model:

Vision pipeline:

Results when compared with GroundTruth

• Use model/ScoreCalculator.py for F1 score calculation
• The results might seem low, but to be honest F1 Score wrt grouth truth is little harsh for this use case.

Adam vs RMSProp Optimer

• Adam (Used for this work as giving better F1 Scores and also lower overfitting)

• RMSProp

TODO

• Optimize pipeline for faster processing (using producer-consumer)
• Optimize transfer learning model
• Better Designing (OOPS Aspect)