Labeling an image without geometry

[alexbw]

I would like to label an image without annotating any geometry in the image. For instance, I might want to know whether an image contains a sunset or not, but not the location of the sunset. So, an ideal workflow would be to be able to hit a hotkey, like "s" to indicate a sunset is present, and then hit "page down" to advance to the next image. The annotation might be stored like:

[
    {
        "class": "image",
        "filename": "image_of_sunset.jpg",
        "annotations": [
            {
                "class": "label",
                "name": "sunset", 
            }
        ] 
    }, 
    {
        "class": "image", 
        "filename": "image_of_computer.jpg",
        "annotations": [
            {
            }
        ] 
    }
]

Is this possible?

[baeuml]

No, right now this is not possible. I always imagined to have some "tag"-item to achieve this, but never got around to implement it, mostly because I never needed it personally.

However, I have heard from other people who wanted something like this as well. As a workaround, I think they usually ended up adding a since rectangle to each image, and then setting individual keys for each of their tags to true/false:

[ { "class": "image", "filename": "image_of_sunset.jpg", "annotations": [ { "class": "rect", "x": ..., "y": ..., "width": ..., "height": ..., "sunset": true, } ] }, { "class": "image", "filename": "image_of_computer.jpg", "annotations": [ { "class": "rect", "x": ..., "y": ..., "width": ..., "height": ..., "sunset": false, } ] } ]

If you use a custom RectItem and set defaultAutoTextKeys to the list of tags (here: ["sunset"]), the current state of tags will automatically be displayed (e.g., see examples/customrectitemconfig.py https://github.com/cvhciKIT/sloth/blob/master/examples/customrectitemconfig.py)

On Mon, Nov 3, 2014 at 9:13 PM, Alex Wiltschko notifications@github.com wrote:

I would like to label an image without annotating any geometry in the image. For instance, I might want to know whether an image contains a sunset or not, but not the location of the sunset. So, an ideal workflow would be to be able to hit a hotkey, like "s" to indicate a sunset is present, and then hit "page down" to advance to the next image. The annotation might be stored like:

[ { "class": "image", "filename": "image_of_sunset.jpg", "annotations": [ { "class": "label", "name": "sunset", } ] }, { "class": "image", "filename": "image_of_computer.jpg", "annotations": [ { } ] } ]

Is this possible?

— Reply to this email directly or view it on GitHub https://github.com/cvhciKIT/sloth/issues/88.

Martin Bäuml Karlsruhe Institute of Technology Institute for Anthropomatics Vincenz-Priessnitz-Straße 3, 76131 Karlsruhe, Germany Email: baeuml@kit.edu Phone: +49 721 608 44735, Fax: +49 721 608 45939

[alexbw]

Thanks, I'll look into this.