ovidijusparsiunas / myvision

Computer vision based ML training data generation tool :rocket:

Description

MyVision is a free online image annotation tool used for generating computer vision based ML training data. It is designed with the user in mind, offering features to speed up the labelling process and help maintain workflows with large datasets.

Features

Draw bounding boxes and polygons to label your objects:

Polygon manipulation is enriched with additional features to edit, remove and add new points:

Supported dataset formats:

Annotating objects can be a difficult task... You can skip all the hard work and use a pre-trained machine learning model to automatically annotate the objects for you. MyVision leverages the popular 'COCO-SSD' model to generate bounding boxes for your images and by operating locally on your browser - retain all data within the privacy of your computer:

You can import existing annotation projects and continue working on them in MyVision. This process can also be used to convert datasets from one format to another:

Languages

MyVision is available in English and Chinese (Mandarin).

Local setup

No setup is required to run this project, open the index.html file and you are all set! However, if you want to make changes or contribute to this repository, please follow the instructions below:

# Requirements: Node version 10+ and NPM version 6+

# Install node dependencies:
$ npm install

# Run the project in watch mode:
$ npm run watch

# All changes should be made in the src directory and observed in publicDev

Citation

@MISC{MyVision,
   author = {Ovidijus Parsiunas},
   title = {{MyVision}},
   howpublished = {\url{https://github.com/OvidijusParsiunas/myvision}},
   year = {2019},
}

myvision's Issues

Change of class numbers

I auto-annotated a sample image using COCO-SSD pretrained model.
The default class numbers are 0,1,2 for dog, bicycle and car.

How can I change the default class numbers ?

question: how to create my own classnames without the default classes?

when i export, i got unwanted classes like: dog, cat etc. how to mannually set my own classes ?

Custom tensorflow model

Can i load a custom trained model to autodetect labels as COCO-SSD?

Thank you

Issue with drag & drop

Sometimes when I select a new image and accidentaly drag and drop, the application tends to freeze.

Loss of detail with very small bounding boxes

Reproduce:

1. Load a 512x512 image and accurately label a very small object with full zoom.
2. Export (VOC XML or Yolo Text)
3. Open a new tab and reimport

Support COCO Segmentation with RLE format

Hello!

Do you have plans on implementing a support for COCO RLE format annotations and not only polygons?

{"segmentation": {"counts": [], "size": []}}

Also if it is not your priorities, any suggestions how would you go to implement such a feature in terms of your apps architecture?

Thank you for all the work till now.

"New Bounding box" tool writes segmentation polygons in incorrect format for COCO JSON

Noticed this when pycocotools started complaining.

Data format page for COCO (section 1 - Object Detection) mentions that for detection tasks, segmentation field in annotation object must be either RLE or [polygon]:

"segmentation": RLE or [polygon],

Looks like a polygon itself is a list of vertices (however, i could not find any specification on that).

If we take a look at the annotated picture and its json annotation file, we see that segmentation field for annotation with id=0 (made with Bounding box tool) is a list of floats, but for annotation with id=1 (made with Polygon tool) is a list of lists.

Here is the zip with json file - the result of COCO JSON import.
annotations.zip

Could you check this out?

myvision can not deploy to server like flask.

appBundle.2404691effe4ecfa2988.js:2 Uncaught TypeError: Cannot read properties of null (reading 'addEventListener')