COCO JSON support for Import/Export in Classification Wizard (Training Samples Manager)

[Louis-VincentGrandMaison]

While training samples can currently be imported/exported via feature classes, adding COCO JSON support would significantly improve interoperability with modern ML and computer vision workflows. COCO JSON is the industry-standard annotation format used across tools like CVAT, Labelbox, and frameworks like TensorFlow and PyTorch.

The proposed workflow would allow:

• Importing COCO JSON files (polygons, points, labels) directly as training samples, avoiding manual conversion for users who annotate in external tools.

• Exporting training samples to COCO JSON for seamless use in external ML pipelines or collaboration with non-geospatial teams.

This would allow for better integration between GIS and computer vision, reduce redundant work, and align ArcGIS Pro with widely adopted AI standards. Could this be considered for a future update?

[PavanYadav]

@Louis-VincentGrandMaison

Thank you for sharing this idea. I will share it with my team and see what we can do. Is this something that you use fairly often? Currently, for object detection, we support Pascal VOC and KITTI rectangles. I may have a script that can convert COCO JSON to Pascal VOC to use them in ArcGIS. If you're interested, I can share it with you.

Thanks again for your feedback and idea!

[Louis-VincentGrandMaison]

Thanks for your reply! While conversion scripts are helpful (and I’d appreciate seeing yours), native COCO JSON support would still provide key advantages:

1. Broader Compatibility: Many newer tools (e.g., Roboflow, SAHI for small object detection) and models (YOLOv8, Detectron2) default to COCO format. Manual conversion adds steps and can lose metadata (e.g., segmentation masks vs. bounding boxes).

2. Complex Annotation Support: COCO handles not just rectangles (like Pascal VOC/KITTI) but also:

   • Precise polygon annotations for irregular features (e.g., tree canopies, water bodies).

   • Keypoints (e.g., for infrastructure inspection).

   • Crowd annotations (dense small objects).

3. Collaboration Efficiency: Teams often standardize on COCO for cross-platform projects. Native support would let ArcGIS Pro users seamlessly share data with non-GIS ML teams.

4. Dataset management: COCO’s structure is more conducive to dataset versioning and dataset management, especially when working in collaborative environments or with automated data pipelines. It allows a clearer separation between image metadata, annotation metadata, and category definitions, which facilitates dataset updates and the integration of new annotation types without disrupting the overall format.

In my mind, COCO’s dominance in computer vision makes this a natural extension to ArcGIS pro's existing capabilities, especially as GIS/ML integration grows.