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
:blue_book: Automatic documentation from sources, for MkDocs.

mkdocstrings Automatic documentation from sources, for MkDocs. Features - Python handler - Requirements - Installation - Quick usage Features Language

1.1k Jan 04, 2023
FxBuzzly - Buzzly.art links do not embed in Discord, this fixes them (rudimentarily)

fxBuzzly Buzzly.art links do not embed in Discord, this fixes them (rudimentaril

Dania Rifki 2 Oct 27, 2022
Some code that takes a pipe-separated input and converts that into a table!

tablemaker A program that takes an input: a | b | c # With comments as well. e | f | g h | i |jk And converts it to a table: ┌───┬───┬────┐ │ a │ b │

CodingSoda 2 Aug 30, 2022
Python document object mapper (load python object from JSON and vice-versa)

lupin is a Python JSON object mapper lupin is meant to help in serializing python objects to JSON and unserializing JSON data to python objects. Insta

Aurélien Amilin 24 Nov 09, 2022
Yet Another MkDocs Parser

yamp Motivation You want to document your project. You make an effort and write docstrings. You try Sphinx. You think it sucks and it's slow -- I did.

Max Halford 10 May 20, 2022
Xanadu Quantum Codebook is an experimental, exercise-based introduction to quantum computing using PennyLane.

Xanadu Quantum Codebook The Xanadu Quantum Codebook is an experimental, exercise-based introduction to quantum computing using PennyLane. This reposit

Xanadu 43 Dec 09, 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
LotteryBuyPredictionWebApp - Lottery Purchase Prediction Model

Lottery Purchase Prediction Model Objective and Goal Predict the lottery type th

Wanxuan Zhang 2 Feb 14, 2022
Modified fork of CPython's ast module that parses `# type:` comments

Typed AST typed_ast is a Python 3 package that provides a Python 2.7 and Python 3 parser similar to the standard ast library. Unlike ast up to Python

Python 217 Dec 06, 2022
Zero configuration Airflow plugin that let you manage your DAG files.

simple-dag-editor SimpleDagEditor is a zero configuration plugin for Apache Airflow. It provides a file managing interface that points to your dag_fol

30 Jul 20, 2022
step by step guide for beginners for getting started with open source

Step-by-Step Guide for beginners for getting started with Open-Source Here The Contribution Begins 💻 If you are a beginner then this repository is fo

Arpit Jain 66 Jan 03, 2023
Data-science-on-gcp - Source code accompanying book: Data Science on the Google Cloud Platform, Valliappa Lakshmanan, O'Reilly 2017

data-science-on-gcp Source code accompanying book: Data Science on the Google Cloud Platform, 2nd Edition Valliappa Lakshmanan O'Reilly, Jan 2022 Bran

Google Cloud Platform 1.2k Dec 28, 2022
An awesome Data Science repository to learn and apply for real world problems.

AWESOME DATA SCIENCE An open source Data Science repository to learn and apply towards solving real world problems. This is a shortcut path to start s

Academic.io 20.3k Jan 09, 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
Data science Python notebooks: Deep learning (TensorFlow, Theano, Caffe, Keras), scikit-learn, Kaggle, big data (Spark, Hadoop MapReduce, HDFS), matplotlib, pandas, NumPy, SciPy, Python essentials, AWS, and various command lines.

Data science Python notebooks: Deep learning (TensorFlow, Theano, Caffe, Keras), scikit-learn, Kaggle, big data (Spark, Hadoop MapReduce, HDFS), matplotlib, pandas, NumPy, SciPy, Python essentials, A

Donne Martin 24.5k Jan 09, 2023
Leetcode Practice

LeetCode Practice Description This is my LeetCode Practice. Visit LeetCode Website for detailed question description. The code in this repository has

Leo Hsieh 75 Dec 27, 2022
Collections of Beautiful Latex Snippets

HandyLatex Collections of Beautiful Latex Snippets Table 👉 Succinct table with bold separation line and gray text %################## Dependencies ##

Xintao 15 Apr 11, 2022
A hack to run custom shell commands when building documentation on Read the Docs.

readthedocs-custom-steps A hack to run custom steps when building documentation on Read the Docs. Important: This module should not be installed outsi

Niklas Rosenstein 5 Feb 22, 2022
NetBox plugin for BGP related objects documentation

Netbox BGP Plugin Netbox plugin for BGP related objects documentation. Compatibility This plugin in compatible with NetBox 2.10 and later. Installatio

Nikolay Yuzefovich 133 Dec 27, 2022
Fastest Git client for Emacs.

EAF Git Client EAF Git is git client application for the Emacs Application Framework. The advantages of EAF Git are: Large log browse: support 1 milli

Emacs Application Framework 31 Dec 02, 2022