CVAT series opening

0. Preface

• It seems that there is no CVAT information on the Internet, so you need to read the official manual carefully.
• I’m going to use it now. I hope to write a series as an internal communication document for the company, and share it with labeling engineers like me who have no money to find someone to label and have to do it themselves.

0.1. Problems with annotation tools used before

• The annotation tool used before is VoTT
• I won’t mention some labeling details here. After all, labeling is physical work, and the tools are almost the same after they are used to them.
• The main problem is that it is not suitable for multi-person collaboration .
  • I used VoTT labeling and 6 people labeling before, and I ran into many, many problems.
    • Cannot import other people's annotation results.
    • Task assignment needs to be processed offline, which is very troublesome.
    • Cannot see the progress of everyone's work.
  • Wait, wait (all kinds of bitter tears)...
• The program is unstable: VoTT is a desktop application and often crashes.

0.2. Why use CVAT?

• The main reason is that it is more suitable for multi-person collaboration .
• Can be used for almost all CV related labeling tasks.
• Has been updated.

0.3. Impressions

• If you are just one person, and slightly mark some data, in fact, VoTT is a very good choice.
• If there is a lot of labeled data and multi-person collaboration is involved, the advantages of CVAT will be reflected.
• VoTT's learning cost is low, installation is simple, and it's quick to get started; CVAT is complicated to install (a colleague took a week to install it, doge), and it has full functions.

0.4. Related Information

• There are very few CVAT related materials on the Internet, where can I find it?

  • First of all, you must read all the official documents, especially the user manual.
  • Second, check in Github Issues .
  • Again, search by keyword on Gitter .
  • Finally, Baidu/Google keyword search.
  • No, just ask questions in gitter, github issue (in fact, the reply is quite fast).

• Official information:

  • Github
  • Online Demo : Official sample, free trial, maximum data volume of each account is 500M, maximum 10 tasks
  • Official user manual translation
  • CVPR 2019 Tutorials，以及slides

• other information

  • Installation reference

1. Follow-up study plan

• CVAT installation of mining pits

  • Various versions of CVAT (v1.1.0/v1.2.0-alpha/develop) have been successfully installed on the company server (Ubuntu Server 16.04), my own computer (Ubuntu 16.04 Desktop), and Alibaba Cloud server (Ubuntu Server 18.04).
  • Stepping on a lot of pits, it should be able to help some rookies like me.

• A simple example of CVAT

  • In just the case of installed CVAT, and how to label a simple bbox example.
  • I hope to be able to use as many functions as possible. Although I will not introduce every function in detail, I will show that there are these functions.

• Introduction to CVAT interface section

  • The notes have been completed, please refer to "Using CVAT (1) Interface and Annotation Function Introduction"
  • This note has little content, it can be regarded as a note from reading the user manual for yourself, mainly because you first understand what functions CVAT supports as a whole.
  • main content:
    • Introduce the CVAT annotation interface.
    • Introduce some basic functions supported by CVAT.

• Detailed user management function

  • The notes have been completed, please refer to "CVAT Use (2) User Management"

  • It is worth mentioning that the current * (2020.12.5) version of CVAT user management has problems, and the development team said that it is developing a new authority management function.

  • The main contents are:

    • What permissions does a super user have and what functions can be implemented.
    • What permissions exist for ordinary users.

• Detailed Explanation of Tasks and Jobs in CVAT

  • The notes have been completed, please refer to "CVAT Use (3) Tasks/Jobs creation and management"

  • The main contents include:

    • Detailed introduction of tasks creation. Including the detailed explanation of labels.
    • Creation of jobs.
    • The tasks/jobs management function provided in CVAT.

• Detailed explanation of pre-training model (semi-automatic labeling)

  • How to implement semi-automatic labeling.
  • What pre-trained models are provided.
  • If we have our own model, how to import and use it.

• Multi-person collaborative annotation exploration

  • What are the different permissions when multiple people collaborate.
  • What are the basic processes of allocation, labeling, and review, and where are they implemented?

• Detailed explanation of the labeling process (including labeling tips)

  • According to the labeling task, the basic labeling process is introduced.
  • Introduce the use of shortcut keys.
  • Introduce some other problems encountered.
  • Including input import form.

• Import and export of annotation results

  • What is the export format of the annotation result, how to export it, and what other options are there.
  • How to import existing annotation data?

2. Organizing and reading official documents

• Installation guide , the installation document. For more related content, please refer to Chapter 3 of this article.

  • The main content includes the basic installation process on Ubuntu 18.04/Win10/Mac.
  • There are also some optional options during the installation process: set proxy, semi-automatic label deployment, stop all containers, set custom domain name/IP, email verification, shared hard drive and other functions.

• User's guide , the user manual

  • This part is the most content, just write what is included, and more information is written in the detailed usage record.
  • Getting Started: An introductory tutorial, how to mark for the first time after installing CVAT, and introduces the semi-automatic mark and search functions.
  • Interface of the annotation tool: A brief introduction to the interface of CVAT, its essence is to introduce the functions of CVAT.
  • Introduced several modes in CVAT, namely shape mode, track mode, attribute annotation mode.
  • Introduced several common labeling methods and basic operating procedures, namely annotation with rectangle by 4 points/polygons/polylines/points/cuboids/tags/polygons
  • Introduced the automatic annotation function, namely Automatic annotation.
  • There are also Shape Group (similar to combination), Filter (filter annotation) functions.
  • All available shortcut keys are introduced. This is very useful, you can read it from beginning to end. During the marking process, shortcut keys can greatly improve efficiency.

• XML annotation format : When downloading annotation results from CVAT, it is likely to be in XML format.

• Django REST API : Some apis are provided after deployment, but I don't care too much.

• Datumaro dataset framework : Data set management tool, seems to have some format conversion, quality check and other functions.

• Command line interface : A command line tool, a python script that provides functions for task management, downloading pictures, and downloading annotations.

• AWS deployment Guide : Precautions for AWS deployment

• FAQ

• Questions : How to ask questions