Street-View-House-Numbers-Detection

The proposed challenge is a street view house numbers detection, which contains two parts:

1. Do bounding box regression to find top, left, width and height of bounding boxes which contain digits in a given image
2. classify the digits of bounding boxes into 10 classes (0-9)

The giving SVHN dataset contains 33402 images for training and 13068 images for testing. This project uses the YOLOv5 pre-trained model to fix this challenge.

Hardware

• Intel(R) Core(TM) i5-9600K CPU @ 3.70GHz
• NVIDIA GeForce RTX 2080 Ti

Environment

• Microsoft win10
• Python 3.7.3
• Pytorch 1.7.0
• CUDA 10.2

Reproducing Submission

To reproduct my submission without retrainig, do the following steps:

1. Installation
2. Data Preparation
3. Set Configuration
4. Download Pretrained Model
5. Training
6. Testing
7. Reference

Install Packages

• install pytorch from https://pytorch.org/get-started/locally/
• install openCV

sudo apt-get install python-opencv

• install dependencies

pip install -r requirements.txt

Data Preparation

Download the given dataset from Google Drive or SVHN Dataset.

data / svhn
  +- train
  |	+- xxx.jpg
  |	+- digitStruct.mat
  +- test
  |	+- yyy.jpg
  +- mat_to_yolo.py
  +- shvn.yaml

And run command python mat_to_yolo.py to create labels for yolo and reorganize the train data structure as below:

- train/
├── 1.png
├── 1.txt
├── 2.png
├── 2.txt
│     .
│     .
│     .
├── 33402.png
└── 33402.txt

Set Configuration

• create svhn.yaml in ./data

# train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
train: data/svhn/train  # 33402 images
val: data/svhn/valid  # 3000 images

# number of classes
nc: 10

# class names
names: ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']

Download Pretrained Model

• https://github.com/ultralytics/yolov5/releases

Training

• train model with pretrained model

python train.py --img 320 --batch 16 --epochs 50 --data svhn.yaml --weights yolov5m.pt

• Using the following script to get more information

python train.py --help

Testing

• detect test data

python detect.py --source data/svhn/test/ --weights runs/train/exp18/weights/best.pt --conf 0.25 --save-txt --save-conf

• Using the following script to get more information

python detect.py --help

• Make Submission: output json format

python combine.py

[{
  "bbox": [[top, left, buttom, right]],
  "score": [confidence],
  "label": [predict_label]
 }, 
 {
  "bbox": [[7, 112, 28, 121], [9, 122, 27, 134],
  "score": [0.674805, 0.713867],
  "label": [1, 0]
 }
]

Reference

• h5py - Quick Start Guide
• YOLOv5