Export labelme json files as geojson files

[aloboa]

In order to be able to use other software packages with polygons generated with labelme, converting LabelMe JSON files into GeoJSON files would be most useful.

I have asked chatgpt to make a json -> geojson conversion:

• labelme json file (with no image):
  carbones_noimage.json

• chatgpt geojson file:
  https://www.dropbox.com/scl/fi/nl5gkd3wykvr4jkw07aei/carbones.geojson?rlkey=n82b3ik39y1e0akd2ugjs4m2s&dl=0
  (cannot upload the file as I get "We don’t support that file type" message)

The difference looks really simple; it is just a matter of replacing these lines at the beginning:

{
  "version": "5.2.1",
  "flags": {},
  "shapes": [
    {
      "label": "charcoal",
      "points": [

with

{
 "type": "FeatureCollection",
 "crs": {
   "type": "name",
   "properties": {
     "name": "LOCAL_CS[\"LabelMe image space\"]"
   }
 },
 "features": [
   {
     "type": "Feature",
     "geometry": {
       "type": "Polygon",
       "coordinates": [

It would be very handy having the geojson format as an option in the "Save as" menu

[wkentaro]

other software packages

what software you have in your mind?

[aloboa]

We often deal with satellite, aerial and drone images, which are georeferenced. Therefore, we would need that polygons
generated with labelme were also georeferenced. geojson format is accepted by most (all?)
GIS software (eg, QGIS). In this way, polygons generated with labelme could also be used in a GIS context.
We also have quite a number of polygons drawn using GIS software that we would like to incorporate to labelme. One way could be to export as geojson and have labelme accepting geojson, and an alternative is to export polygons as label raster images that are accepted by labelme.

[minewilliam]

I do not believe it is Labelme's responsibility to handle multiple file formats; But rather to provide a (json) format that is suitable for the end-user to ingest and transform the data themselves.

By introducing a "Save as" feature, it opens the door to users wanting to (re)-import their annotation data. While this change would be quite trivial to implement, I do not think it is sustainable to support parsing & serializing across multiple (possibly custom) storage formats.

I think that a simple, well defined, code interface for custom data adapters would be the ideal solution here.

[aloboa]

I disagree.
It is not a matter of providing many different formats, but at least one (geojson) that can be used with georeferenced imagery (satellite, aerial, drone). Closing the door to geojson you close the door to a significant number of applications. Labelling in the remote sensing community has been common ("photointerpretation") long before AI had even been named.

[kancheng]

@aloboa @minewilliam @wkentaro

I think it might be a good idea to develop a separate, dedicated project for converting Labelme JSON annotations into different target formats, and maintain it as an external tool.

I have previously worked on a related data-conversion project here:
https://github.com/kancheng/data-conversion

This project was originally developed to meet practical image segmentation requirements in an industrial environment.

From my perspective, keeping the core design of Labelme lightweight and avoiding excessive complexity is also important.