300 Detection Limit in Downloaded Model Weights

🤝 Community Help
,
1

I trained a YOLO26 instance segmentation model. My images contain 500+ objects, but I’m consistently getting only ~300 detections.

What I’ve tried:

  1. Roboflow Hosted API: Different confidence/IOU thresholds - always caps at 300
  2. Purchased model weights ($99): Downloaded to self-host with unlimited detections
  3. Roboflow Inference SDK: Deployed using the open-source inference library - still 300 limit
  4. Ultralytics: Built custom inference server with Ultralytics 8.4.6 using the downloaded weights
  5. Model inspection: Found max_det: 300 hardcoded in the Segment26 detection head of the model architecture

Questions:

  1. Is there a 300 detection limit built into models trained on Roboflow, even when self-hosting?
  2. If yes, how can I remove this limit from the downloaded weights?
  3. Can models be trained/retrained with a higher max_det value (e.g., 2000)?

I purchased the weights expecting to bypass the hosted API’s limitations, but the limit appears to be in the model architecture itself. Has anyone encountered this and found a solution?

Project Type: Instance Segmentation
Operating System & Browser: macOS / Chrome
Do you grant Roboflow Support permission to access your Workspace for troubleshooting? (Yes/No): Yes

2

Not sure if you saw this post from a year ago, but it indicates inference-sdk should work. Maybe check out the provided code in there and see if anything is different compared to what you tried.

3

Thanks, I’ve already looked at that and tried it out, but I’m still hitting the same 300 detection limit.

I’ve tried a few different setups at this point. The Roboflow hosted API always caps at 300 detections no matter how I change confidence or IOU. I also deployed inference-sdk myself on Google Cloud and explicitly set max_detections=2000, but the output still stops at 300. I even bought the model weights and ran inference locally, same result there too.

What’s throwing me off is that this limit seems to persist everywhere. Changing max_detections doesn’t appear to do anything, so I’m not sure if this is an inference-sdk thing, something baked into the trained model, or just a hard limit in the YOLO architecture.

For the Roboflow team:

  • Is there a max_det value fixed during training that defaults to 300?
  • When using inference-sdk with max_detections=2000, should it actually be possible to exceed 300, or is the model still capped?
  • Can models be retrained with a higher detection limit, or is this fundamentally a YOLO constraint?
  • Are purchased weights any different from the API models when it comes to detection limits?
4

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.