level2-data-annotation_cv-level2-cv-15 created by GitHub Classroom

Overview

[AI Tech 3기 Level2 P Stage] 글자 검출 대회

image

팀원 소개

김규리_T3016 박정현_T3094 석진혁_T3109 손정균_T3111 이현진_T3174 임종현_T3182

Overview

OCR (Optimal Character Recognition) 기술은 사람이 직접 쓰거나 이미지 속에 있는 문자를 얻은 다음 이를 컴퓨터가 인식할 수 있도록 하는 기술로, 컴퓨터 비전 분야에서 현재 널리 쓰이는 대표적인 기술 중 하나입니다.

OCR task는 글자 검출 (text detection), 글자 인식 (text recognition), 정렬기 (Serializer) 등의 모듈로 이루어져 있는데 본 대회는 글자 검출 (text detection)만을 해결하게 됩니다.

데이터를 구성하고 활용하는 방법에 집중하는 것을 장려하는 취지에서, 제공되는 베이스 코드 중 모델과 관련한 부분을 변경하는 것이 금지되어 있습니다. 데이터 수집과 preprocessing, data augmentation 그리고 optimizer, learning scheduler 등 최적화 방식을 변경할 수 있습니다.

  • Input : 글자가 포함된 전체 이미지
  • Output : bbox 좌표가 포함된 UFO Format

평가방법

  • DetEval

    이미지 레벨에서 정답 박스가 여러개 존재하고, 예측한 박스가 여러개가 있을 경우, 박스끼리의 다중 매칭을 허용하여 점수를 주는 평가방법 중 하나 입니다

    1. 모든 정답/예측박스들에 대해서 Area Recall, Area Precision을 미리 계산해냅니다.

    2. 모든 정답 박스와 예측 박스를 순회하면서, 매칭이 되었는지 판단하여 박스 레벨로 정답 여부를 측정합니다.

    3. 모든 이미지에 대하여 Recall, Precision을 구한 이후, 최종 F1-Score은 모든 이미지 레벨에서 측정 값의 평균으로 측정됩니다.

      image

Final Score  🏅

  • Public : f1 0.6897 → Private f1 : 0.6751
  • Public : 11위/19팀 → Private : 9위/19팀

image

Archive contents

template
├──code
│  ├──augmentation.py
│  ├──convert_mlt.py
│  ├──dataset.py
│  ├──deteval.py
│  ├──east_dataset.py
│  ├──inference.py
│  ├──loss.py
│  ├──model.py
│  └──train.py
└──input
   └──ICDAR2017_Korean
		  └──data
			  	├──images
		      └──ufo
			        ├──train.json
							└──val.json

Dataset

  • ICDAR MLT17 Korean : 536 images ⊆ ICDAR MLT17 : 7,200 images

  • ICDAR MLT19 : 10,000 images

  • ICAR ArT : 5,603 images

Experiment

Results

dataset 데이터 수 LB score (public→private) Recall Precision
01 ICDAR17_Korean 536 0.4469 → 0.4732 0.3580 → 0.3803 0.5944 → 0.6264
02 Camper (폴리곤 수정 전) 1288 0.4543 → 0.5282 0.3627 → 0.4349 0.6077 → 0.6727
03 Camper (폴리곤 수정 후) 1288 0.4644 → 0.5298 0.3491 → 0.4294 0.6936 → 0.6913
04 ICDAR17_Korean + Camper 1824 0.4447 → 0.5155 0.3471 → 0.4129 0.6183 → 0.6858
05 ICDAR17(859) 859 0.5435 → 0.5704 0.4510 → 0.4713 0.6837 → 0.7222
06 ICDAR17_MLT 7200 0.6749 → 0.6751 0.5877 → 0.5887 0.7927 → 0.7912
07 ICDAR19+ArT 약 15000 0.6344 → 0.6404 0.5489 → 0.5607 0.7514 → 0.7465

Requirements

pip install -r requirements.txt

UFO Format으로 변환

python convert_mlt.py

SRC_DATASET_DIR = {변환 전 data 경로}

DST_DATASET_DIR = {변환 된 data 경로}

UFO Format ****

File Name
    ├── img_h
    ├── img_w
    └── words
        ├── points
        ├── transcription
        ├── language
        ├── illegibillity
        ├── orientation
        └── word_tags

Train.py

python train.py --data_dir {train data path} --val_data_dir {val data path} --name {wandb run name} --exp_name {model name
Plotting and analysis tools for ARTIS simulations

Artistools Artistools is collection of plotting, analysis, and file format conversion tools for the ARTIS radiative transfer code. Installation First

ARTIS Monte Carlo Radiative Transfer 8 Nov 07, 2022
🌱 Complete API wrapper of Seedr.cc

Python API Wrapper of Seedr.cc Table of Contents Installation How I got the API endpoints? Start Guide Getting Token Logging with Username and Passwor

Hemanta Pokharel 43 Dec 26, 2022
A python package to avoid writing and maintaining duplicated python docstrings.

docstring-inheritance is a python package to avoid writing and maintaining duplicated python docstrings.

Antoine Dechaume 15 Dec 07, 2022
sphinx builder that outputs markdown files.

sphinx-markdown-builder sphinx builder that outputs markdown files Please ★ this repo if you found it useful ★ ★ ★ If you want frontmatter support ple

Clay Risser 144 Jan 06, 2023
Comprehensive Python Cheatsheet

Comprehensive Python Cheatsheet Download text file, Buy PDF, Fork me on GitHub or Check out FAQ. Contents 1. Collections: List, Dictionary, Set, Tuple

Jefferson 1 Jan 23, 2022
A clean customizable documentation theme for Sphinx

A clean customizable documentation theme for Sphinx

Pradyun Gedam 1.5k Jan 06, 2023
Gtech μLearn Sample_bot

Ser_bot Gtech μLearn Sample_bot Do Greet a newly joined member in a channel (random message) While adding a reaction to a message send a message to a

Jerin Paul 1 Jan 19, 2022
Main repository for the Sphinx documentation builder

Sphinx Sphinx is a tool that makes it easy to create intelligent and beautiful documentation for Python projects (or other documents consisting of mul

5.1k Jan 02, 2023
Code and pre-trained models for "ReasonBert: Pre-trained to Reason with Distant Supervision", EMNLP'2021

ReasonBERT Code and pre-trained models for ReasonBert: Pre-trained to Reason with Distant Supervision, EMNLP'2021 Pretrained Models The pretrained mod

SunLab-OSU 29 Dec 19, 2022
A curated list of python programming language blogs

Python Blogs A curated list of python programming language blogs Contribute Companies/Organization # A B C D E F G H I J K L M N O P Q R S T U V W X Y

Rizky D. Onto 48 Nov 15, 2022
python wrapper for simple-icons

simpleicons Use a wide-range of icons derived from the simple-icons repo in python. Go to their website for a full list of icons. The slug version mus

Sachin Raja 14 Nov 07, 2022
[Unofficial] Python PEP in EPUB format

PEPs in EPUB format This is a unofficial repository where I stock all valid PEPs in the EPUB format. Repository Cloning git clone --recursive Mickaël Schoentgen 9 Oct 12, 2022

k3heap is a binary min heap implemented with reference

k3heap k3heap is a binary min heap implemented with reference k3heap is a component of pykit3 project: a python3 toolkit set. In this module RefHeap i

pykit3 1 Nov 13, 2021
OpenAPI Spec validator

OpenAPI Spec validator About OpenAPI Spec Validator is a Python library that validates OpenAPI Specs against the OpenAPI 2.0 (aka Swagger) and OpenAPI

A 241 Jan 05, 2023
Generate YARA rules for OOXML documents using ZIP local header metadata.

apooxml Generate YARA rules for OOXML documents using ZIP local header metadata. To learn more about this tool and the methodology behind it, check ou

MANDIANT 34 Jan 26, 2022
Reproducible Data Science at Scale!

Pachyderm: The Data Foundation for Machine Learning Pachyderm provides the data layer that allows machine learning teams to productionize and scale th

Pachyderm 5.7k Dec 29, 2022
advance python series: Data Classes, OOPs, python

Working With Pydantic - Built-in Data Process ========================== Normal way to process data (reading json file): the normal princiople, it's f

Phung Hưng Binh 1 Nov 08, 2021
Fully typesafe, Rust-like Result and Option types for Python

safetywrap Fully typesafe, Rust-inspired wrapper types for Python values Summary This library provides two main wrappers: Result and Option. These typ

Matthew Planchard 32 Dec 25, 2022
Run `black` on python code blocks in documentation files

blacken-docs Run black on python code blocks in documentation files. install pip install blacken-docs usage blacken-docs provides a single executable

Anthony Sottile 460 Dec 23, 2022
Żmija is a simple universal code generation tool.

Żmija Żmija is a simple universal code generation tool. It is intended to be used as a means to generate code that is both efficient and easily mainta

Adrian Samoticha 2 Nov 23, 2021