Project Goal: Provide convenient tool for management of Computer Vision datasets within projects providing numerous experiments with visual images.
Solved Problem: Computer Vision projects providing numerous experiments with the image data usually needs collaborative sharing of these data and supporting wide range of datasets formats.
Among popular image dataset annotation tools github.com/opencv/cvat has so far a largest number of supported formats (Pascal VOC, YOLO, COCO, etc., see the table below).
Computer Vision Annotation Tool (CVAT)/a> open-sourced by Intel just lucking the support of such clouds like AWS. CVAT team has shared a post saying that the AWS "issue in backlog at the moment till we have resources to cover it".
The Model Garden tool is an addition to CVAT providing the following functionality:
- store in S3, reuse and modify image datasets using CVAT tool
- collaborative usage of datasets through web interface
- prevent labeling tool crash consequences
NOTE: The currently supported version of CVAT backend API is 0.6.1.
| Usage Order | Tool | Publisher | Web | Cloud | Pascal VOC | YOLO | COCO | MASK | TFRecord | MOT |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | github.com/tzutalin/labelImg | private | N | N | Y | Y | N | N | N | N |
| 2 | github.com/opencv/cvat | Intel | Y | N | Y | Y | Y | Y | Y | Y |
| 3 | github.com/microsoft/VoTT | Microsoft | N | Y | Y | N | N | N | N | N |
Model Garden was started as EPAM Systems internal initiative to support EPAM Computer Vision teams (e.g. Vudoku Accelerator).
The project is open sourced with the support of epam.github.io.
🖌️ Material Design: Intuitive UI based on the world's most widespread design language.
🏃 Single Page Application: Fast, responsive ux to get what you need done without waiting for full-screen refreshes.
🐍 Python Django and Postgres
🏷 AWS S3 DataSet Gallery
Model Garden supports case when with only one data scientist works on the project (does labeling and ML training). This can be useful for experimental projects.
Model Garden supports the case of Massive parallel labeling when a manager has a lot of images and a list of labels. In this case, the manager can upload DataSet, create a list of labels, and assign certain DataSet parts to different labelers, and next control their work.
Material Garden is a mediator between CVat (one of the most popular open-source annotation tool for computer vision) and Amazon S3 (object storage service).
Status Worker is part of Material Garden. This worker checks event updates from CVAT asynchronously.
The example of Continuous Deployment to the cloud container registry as well as the cloud update is set with help of GitLab CI/CD .gitlab-ci.yml file.
.gitlab-ci.yml file needs the following GitLab CI/CD nvironment variables set:
DEV_AWS_ACCESS_KEY_ID=<ABCDEFGHIJKLMNOPQRST>*
DEV_AWS_SECRET_ACCESS_KEY=<abcdefghijklmnopqrstuvwxyz0123456789-+/>*
DEV_BACKEND_ECR_URI=123456789000.dkr.ecr.eu-central-1.amazonaws.com/model_garden_backend
DEV_FRONTEND_ECR_URI=123456789000.dkr.ecr.eu-central-1.amazonaws.com/model_garden_frontend
PROD_AWS_ACCESS_KEY_ID='<ABCDEFGHIJKLMNOPQRST>*'
PROD_AWS_SECRET_ACCESS_KEY=<abcdefghijklmnopqrstuvwxyz0123456789-+/>*'
PROD_BACKEND_ECR_URI=123456789000.dkr.ecr.eu-central-1.amazonaws.com/model_garden_backend
PROD_FRONTEND_ECR_URI=123456789000.dkr.ecr.eu-central-1.amazonaws.com/model_garden_frontend
RELEASE_AWS_ACCESS_KEY_ID='<ABCDEFGHIJKLMNOPQRST>*'
RELEASE_AWS_SECRET_ACCESS_KEY=<abcdefghijklmnopqrstuvwxyz0123456789-+/>*'
RELEASE_BACKEND_ECR_URI=123456789000.dkr.ecr.eu-central-1.amazonaws.com/model_garden_backend
RELEASE_FRONTEND_ECR_UR=123456789000.dkr.ecr.eu-central-1.amazonaws.com/model_garden_frontend
- CVAT Installation Guide
- See installation steps for backend developers in <model_garden_root>/backend/README.md
- See installation steps for frontend developers in <model_garden_root>/frontend/README.md
See all the details in the <model_garden_root>/deploy/README.md.
If CI/CD is set via .gitlab-ci.yml the build pipeline is started automatically after commits to master and develop branches