HOMEWORK 2 - Street View House Number Object Detection

This model is for object detection for the Street Veiw House Numbers dataset. This repo is based on this repo.

Hardware

• Ubuntu 18.04.5 LTS
• Intel® Xeon® Silver 4210 CPU @ 2.20GHz
• NVIDIA GeForce RTX 2080 Ti

Reproduce Submission

To reproduce my submission without training, do the following:

1. Installation
2. Data Preparation
3. Inference

Installation

Install all the requirments.

pip install pycocotools numpy opencv-python tqdm pyymal webcolors torch torchvision

Data Preparation

The data should be placed as follows:

repo
  +- train
  |  +- 1.png
  |  +- 2.png
  |  +- ...
  |
  +- val
  |  +- 32403.png
  |  +- 32404.png
  |  +- ...
  |
  +- test
  |  +- 1.png
  |  +- 2.png
  |  +- ...
  |
  +- annotation_train.csv
  +- annotation_val.csv
  +- train.py
  +- infer.py
  +- weights.pth   (needed for inference)
  +- numbers.yml   (needed for training)
  |  ...

Where train folder contains all the training images, val folder contains all the validation images, and test folder contains all test images. Both annotation_train.csv and annotaion_val.csv should contain the file name, corresponding labels and bounding boxes of each image in train and val folder. Please check annotaion_train.csv to see the expected format.

Training

To train, please download the pretrained weight here and put it beside train.py. Simply run train.py. The weights should be saved in 'logs/numbers' folder. There will be several file saving weights at different training process. The batch_size is set to be 3. Make it smaller in numbers.yml if memory is not sufficent.

Inference

for inference, please download the weights file here and put it beside infer.py. Simply run infer.py and predictions.json containing images file names and their corresponding predictions will be created.

Citation

Yet Anothor EfficientDet Pytorch