SVHN-detection

This repository gathers the code for object detection on SVHN image from the in-class CodaLab competition.

We use YOLOv5, an open source object detection architectures and models based on PyTorch, to train our model. In this competition, we use YOLOv5m, YOLOv5l these two models and analysis the results.

Reproducing Submission

We need to do some pre-preparation for training and testing on our custom dataset.

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

1. Requirement

2. Repository Structure

3. Inference

   or

Just use Inference.ipynb in Google Colab.

Hardware

Ubuntu 18.04.5 LTS

Intel® Core™ i7-3770 CPU @ 3.40GHz × 8

GeForce GTX 1080/PCIe/SSE2

Speed.ipynb: Google Colab

Requirement

All requirements should be detailed in requirements.txt.

$ conda create -n YOLOv5 python=3.7 -y
$ conda activate YOLOv5
$ cd SVHN-object-detection
$ conda install pytorch=1.10.0 torchvision=0.11.1 -c pytorch
$ pip install -r requirements.txt

Official images can be downloaded from CodaLab competition

Repository Structure

The repository structure is:

SVHN-object-detection(root)
  +-- data                       # all files and script used in the program 
  |   +-- hyps
  |   +-- scripts
  |   +-- svhn
      |   +-- Annotations        # training set .xml
          |   +-- 1.xml   
          |   +-- 2.xml   
          |   +-- 3.xml   
          |   +-- ......
      |   +-- test               # testing set .png  
          |   +-- 117.png  
          |   +-- 162.png 
          |   +-- 196.png  
          |   +-- ......
      |   +-- train              # training set .png /.txt
          |   +-- 1.png   
          |   +-- 1.txt   
          |   +-- 2.png 
          |   +-- 2.txt 
          |   +-- ......
      |   +-- trainval           # training and validation sets .png /.txt
          |   +-- 1.png   
          |   +-- 1.txt   
          |   +-- 2.png 
          |   +-- 2.txt 
          |   +-- ......
      |   +-- val                # validation set .png /.txt
          |   +-- 22599.png   
          |   +-- 22599.txt   
          |   +-- 22600.png 
          |   +-- 22600.txt 
          |   +-- ......
      |   +-- digitStruct.mat
      |   +-- json_to_VOC.py
      |   +-- mat_to_json.py     
      |   +-- MatTransform.json  
      |   +-- VOC_to_yolo.py
      |   +-- train_val_split.py
  |   +-- svhn.yaml
  +-- models
  +-- utils
  +-- runs
  |   +-- detect                 # run detect.py and testing .png /.txt save here.
      |   +-- labels
          |   +-- 1.txt  
          |   +-- 2.txt  
          |   +-- ......
      |   +-- 1.png
      |   +-- 2.png
      |   +-- ......
  |   +-- train                  # run train.py and training info save here
  +-- detect.py                  # model prediction
  +-- inference.py               # reproduce my submission file
  +-- train.py                   # for training model
  +-- hubconf.py
  +-- Speed.ipynb                # for testing detector speed
  +-- requirements.txt           # txt file for establishing the environment

Dataset setting

You can use mat_to_json.py to parse 'digitStruct.mat' file to .json file and then use json_to_VOC.py transfer to VOC format. This can use in MMDetection.

Next, using VOC_to_yolo.py to make training set .txt (about info) for YOLOv5.

$ python data/svhn/mat_to_json.py
$ python data/svhn/json_to_VOC.py
$ python data/svhn/VOC_to_yolo.py

Finally, using train_val_split.py to split data, and it will move training and validation data to train and val folder respectively. The trainval folder will be empty.

$ python data/svhn/train_val_split.py

Training

Before training, you need to split training and validation sets. Next, modifying train and val path in data/svhn.yaml.

train: data/svhn/train  # 26722 images
val: data/svhn/val  # 6680 images
nc: 10  # class number
names: ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']  # class names

To train the model, run this command:

$ python train.py --epochs 100 --data svhn.yaml --weights yolov5m.pt

Trained model will be saved as runs/train/exp/weights/best.pt

All mAP and speed of experiments will be written in Results.

Inference

Use Inference.ipynb in Google Colab.

or

Please download this model if you want to reproduce my submission file, and run above codes.

To reproduce my submission file or test the model you trained, run:

$ python detect.py --source data/svhn/test/ --weights runs/train/exp/weights/best.pt --save-conf --save-txt 
$ python inference.py --txt runs/detect/exp/labels/ --data data/svhn/test/

Note the path of model and testing data. --source {testing set path} --weight {your model path}, --data {testing set path}

Prediction file will be saved as root/answer.json

Results

Our model achieves the following performance on :

	YOLOv5m	YOLOv5l	YOLOv5m	YOLOv5m
mAP	0.410383	0.412982	0.409623	0.410516
Speed	-	0.081104	-	0.066749
Note	img 320 epoch100	img 320 epoch100	img 320 epoch100, change train and val sets	default, epoch100

Reference

[1] YOLOv5

[2] YOLO format

[3] VOC format