I’m attempting to parse the output of a yolo11 model created via export to onnx format. When looking at the model in netron, the output looks like this: tensor: float32[1,84,8400]. When looking at a yolo4 model graph is more complex and there are multiple ‘Identity outputs’ ie: 3.
Why is Yolo8 and Yolo11, maybe others onnx model look so different than the yolo4 and how can I properly parse the yolo11 onnx output model into the correct vector space?
You can check the examples here
You can also use nms=True for simpler post-processing.
https://www.reddit.com/r/Ultralytics/comments/1iqp88o/comment/md1w6yv
think that might have fixed the output format. Thank you.
Hello Kevin_PerAcc,
Great to hear the output format issue seems resolved.
Regarding your question about the differences between YOLOv4, YOLOv8, and YOLO11 ONNX model structures, YOLOv8 and YOLO11 models have a simplified architecture that often results in a single output tensor in the ONNX format, as you observed with float32[1,84,8400]. YOLOv4, being implemented with a different architecture, typically has a more complex graph with multiple output tensors.
To parse the YOLO11 ONNX output, you can refer to the Predict mode documentation, and you can find more detailed information about ONNX export in the Export mode documentation. The structure [1, 84, 8400] generally corresponds to the batch size, the number of bounding box attributes (including class scores, and box coordinates), and the number of anchor points, respectively. The export documentation shows all available export formats and their respective arguments. You may also find this information in the Ultralytics ONNX integration documentation.
If you have further questions about specifics, feel free to ask! The YOLO community and the Ultralytics team are here to support you.