ManishSahu53 / Appsilon_challenge

Appsilon_challenge

Overview

1. Create training dataset

Since dataset was not given, I used google image search as source of dataset. I search Cary type in google image search and downloaded ~300 images of each category. I used Firefox addon to download all images at bulk.

2. Deep Neural Network Architecture

Since object dectection and recognition is standard problem and SOTA architecture exist. So I prefer to use YOLOv3 architecture for object detection and recognition.

3. Creating Labels

I used pre-trained YOLOv3 architecture to generate bounding boxes around cars and assigned with suitable class there after. This eliminated my time to manually drawing bounding boxes

4. Training Network

I used 6 six categories with ~1800 labeled bounding boxes datasets.

5. Training Hardware

Training YOlOv3 is GPU extensive task. I don't have a GPU pc so I used google's Colab to train my network. It comes with 12 GB of GPU with latest Nvidia GPU.

Trained model

1. Weights - https://drive.google.com/file/d/1YKvLkQjYgF7rlaejk07DhBtQZE9rhpUL/view?usp=sharing
2. Results - https://drive.google.com/drive/folders/1rfAqpni8vORZtWAbQNNnxlTg5GC11V_D?usp=sharing
3. Yolo Git - https://github.com/ManishSahu53/YoloV3
4. Yolo-pre-trained Weights - wget https://pjreddie.com/media/files/darknet53.conv.74
5. Colab - https://colab.research.google.com/drive/1wnnWwtzdP-FN6rlSr5LbjQ_UuU2dvcRQ

Process Overview

How to Run

1. Clone this repo and Yolo Git - https://github.com/ManishSahu53/YoloV3
2. Download weights from https://drive.google.com/file/d/1YKvLkQjYgF7rlaejk07DhBtQZE9rhpUL/view?usp=sharing as paste it in model_data folder as Cars.h5
3. Copy car_classes.txt present in this repo and paste it in model_data folder present in yolo repo

python test.py --image path_folder_image --model model_data/Cars.h5 --classes model_data/car_classes.txt --output path_folder_output