I’m having a problem with my project, which focuses on Instance Segmentation of intertidal species.
In this project, I have 11 models, and the ones with the highest performance are v11 and v12. However, after creating these two models, I noticed a new issue: when using Auto Label on photos that don’t have annotations, the models are identifying the species as if it were Object Detection (squares) instead of Instance Segmentation.
Previously, this issue didn’t occur when using Auto Label — the segmentation masks were generated correctly. I even tested it again using an older model that worked fine before, but now it also behaves like Object Detection.
I’ve already confirmed that my project type is Instance Segmentation, and that my models are also Instance Segmentation models. To troubleshoot, I duplicated the project and created a new version trained on v12, but the results are still the same — Auto Label only shows Object Detection results instead of segmentation.
I’m not sure why this is happening or what changed, since everything used to work properly before.
Describe your question/issue here! (delete this when you post)
May I ask if you could post how you are inferring the model? (i.e. if you are running locally with inference can you please post the script you are executing with masked API key, or if you are making requests to the inference server can you post the endpoint you are POSTing against)
I’m not sure if it is this but, I’m using the Auto Label feature directly inside Roboflow (not a local script).
Specifically, I select my trained Instance Segmentation model (v12) and use Auto Label on unannotated images within the Roboflow web interface. I’m not calling the inference API manually or using a Python script — all predictions are made through the built-in Auto Label tool.
Previously, this worked perfectly (it generated segmentation masks). However, now the same models only produce object detection-style bounding boxes instead of segmentation masks.
I tried on my test projects and auto label produces masks when segmentation model is selected, can you confirm the model selected has (instance segmentation) at the end of model name when you select the project from the drop-down? This is what I see when I’m enabling label assist on my project:
Once your project is selected, are you selecting all classes? And finally, are all detections appearing as object-detection style, or maybe some portion is appearing as instance segmentation?
Yes, I can confirm that the model I’m selecting does have “instance segmentation” at the end of the name; when enabling Auto Label, I’m selecting all classes (picture); and that all detections are appearing as object detection-style bounding boxes — there are no segmentation masks being produced at all.
I’ve also tested older models that used to generate segmentation masks correctly, and they now show the same behavior — only bounding boxes.
It seems like the Auto Label feature is treating all models as object detection models, even when they’re instance segmentation ones.
Hello. Just to update, this error doesn’t happen anymore. I don’t know the reason, it just started doing Auto label correctly again. Thanks for correcting this error.
