I would also like to raise a broader concern regarding platform reliability for custom computer vision workflows, especially for architectural/floor plan datasets.
Over the last few days I have noticed multiple community posts related to:
-
training failures
-
inference returning no detections
-
dataset sync inconsistencies
-
image upload issues
-
segmentation/export problems
In my own case, both pretrained workflows and custom-trained models frequently return no detections during testing, even after carefully following official tutorials and verifying annotations/configurations.
Because floor plan detection already requires significant time for annotation and training, it becomes difficult to determine whether the issue is:
-
dataset quality,
-
model limitations,
-
incorrect workflow,
or platform-side instability.
It would be very helpful if Roboflow could provide:
-
clearer debugging/error logs,
-
inference confidence diagnostics,
-
architectural/floor-plan-specific guidance,
-
segmentation recommendations,
-
and better visibility into platform-side outages/issues.
I really like the platform and want to continue using it for AI-assisted BIM/floor plan workflows, but currently it is difficult to confidently validate whether the models are actually learning correctly.
Thank you.