Image annotation tool for TuriCreate

Hi there!


I'm looking for some image annotation tool that could produce TuriCreate compatible annotations in JSON format. Or maybe there are some tools to convert annotations (saved in some other formats, like XML) into the JSON format compatible with TurCreate ?


Thanks in advance

Marek

Accepted Reply

Hi,


https://github.com/CivilNet/Gemfield/blob/master/src/python/pascal_voc_xml2json/pascal_voc_xml2json.py

I think it's what you need. It can convert the Pascal Voc Xml to CoCo jason format.

Replies

May be you should move the question to ML forum, you'll probably find better answers there.


The format for annotaions seem pretty straightforward, as shown in WWDC https://developer.apple.com/videos/play/wwdc2018/712/ @ 7'56

[

{

"label" : "coffee", "coordinates" : {

"x" : 387,

"y" : 660,

"height" : 550,

"width" : 814,

}

},

,{

"label" : "croissant", “coordinates" : {

"x" : 800,

"y" : 630,

"height" : 373,

“width" : 812,

}

}

]

sorry, my bad, I was intened to put it into ML section 😊 . It's true that the format for annotations is straightforward. But what if you got more images and/or objects to mark at those images (more than one coffe and croissant 😉). How will you manage to mark objects on hundrets of images? That's why I was asking about some tool to manual marking objects at images


BTW. Is there possibility to move this question to ML section ? 😉

Don't be sorry.


Does this provide the answer ?


I've seen in WWDC that annotations are imported as csv file


// Load annotations & images



annotations = turicreate.SFrame(“annotations.csv”)



images = turicreate.load_images(“training_images”)


What I understand, even if I did not practice, is that you should have 2 files :

- images

- annotations.csv that you have to create

and then join the 2


So the tool is probably simple to write as it would:

- get the name of the image

- let you mark the area

- add the rect to the json,

- add the label


But there seem not to exist yet convenient tool : h ttps://github.com/apple/turicreate/issues/828


And a feature request has alreadu been filed.


In the demo they demonstrate Jupiter Notebook: could that help you ?


To move to ML section, use the Actions pop up button at the top right of the window

Thanks Claude31, but maybe there is some 3rd party solutions for that problem? There are tons of tools like that (https://en.wikipedia.org/wiki/List_of_manual_image_annotation_tools), but not for TuriCreate 😟.

There is an inexpensive tool on the Mac App Store called RectLabel. I don't know if it directly exports a format that Turi can read, but it should be straightforward enough to write a conversion script.

Hi,


https://github.com/CivilNet/Gemfield/blob/master/src/python/pascal_voc_xml2json/pascal_voc_xml2json.py

I think it's what you need. It can convert the Pascal Voc Xml to CoCo jason format.

Hi tao.xi!


It looks interesting. I will check that. Thanks!

I've been struggling with some tools myself, but finally landed on a workflow that works for my team and I.
I'm using RectLabel available on MacOS -- it supports directly exporting into Turi and they provide an additional script that can convert the output CSV into an SFrame. Otherwise, you can load the CSV directly into Turi and have Turi spit out the SFrame for you. It's worked very well for me. The only downsides to RectLabel so far for me is that it is only available on MacOS and does not support team collaboration (these may or may not be an issue depending on your workflow)
I've also recently been dabbling with a less-robust but very efficient app on the iOS App store (iPhone and iPad) called Labelocity. It's simple and can export in CSV for both Turi and Tensorflow. This does support collaboration but at a monthly/annual fee. They don't have a desktop version available, but on iPad it seems to work fairly well. The developer also seems pretty responsive to feedback and such as the tool grows.


I've had minor success with LabelIMG -- It's one of the more well-known solutions if you follow various forums and tutorials online for object detection, but I've had one too many glitches with it that made decide to stop using it altogether. Others seem to have good experience with it though -- so your mileage may vary. It's a free solution and multiplatform -- so that's something. 🙂


I'm sure these are all temporary solutions though until Apple provides a tool for us at next year's WWDC to make image annotations easy and import directly into CreateML without any hassle and conversion. Fingers crossed!

Hello, Try this. It works with both Turi Create and Create ML. It runs on iOS devices and does all the necessary conversions. Touch Annotator by Intersection Bioscientifc Software Inc. https://apps.apple.com/ca/app/touch-annotator/id1478864595

I recently saw this query. I know it is old and tech has improved a lot now. For newby in the space I thought I should share that recently I used a tool called https://www.labellerr.com that support most of the output format in annotation output.