YOLOv8 .pt models fail during inference due to unsupported output formats

[YUVAN0907]

Issue: YOLOv8 .pt models fail during inference in UI

Summary

While testing YOLO-based models in the PerceptionMetrics UI, I ran into runtime errors when using native YOLOv8 .pt models.

The model loads successfully, but inference fails when an image is uploaded.

---

What happens

I observed two types of errors:

1.

'Tensor' object has no attribute 'boxes'

2.

too many values to unpack (expected 2)

---

Steps to reproduce

1. Run the Streamlit UI
2. Select Model Format → YOLO
3. Load a YOLOv8 .pt model (e.g. yolov8n.pt)
4. Upload any image
5. Run inference

The model loads fine, but inference crashes with one of the errors above.

---

Why this happens

It looks like postprocess_detection() assumes the model output is always a torch.Tensor.

However, YOLOv8 models (Ultralytics) can return:

• a Results object (with .boxes)
• a list or tuple
• or a raw tensor

Right now, only the tensor case is handled, so other formats cause crashes.

---

Expected behavior

• Inference should work for YOLOv8 .pt models
• Different output formats should be handled safely
• Bounding boxes should be correctly processed without errors

---

Suggested fix

Update postprocess_detection() to:

• handle Results objects (output.boxes)
• unwrap list/tuple outputs
• keep existing tensor logic unchanged

---

Files involved

• perceptionmetrics/models/utils/yolo.py
• perceptionmetrics/models/torch_detection.py

---

Notes

• This issue blocks inference for native YOLOv8 models
• TorchScript models seem to work fine
• Fixing this would improve compatibility with Ultralytics models

---

Happy to help with a PR if this approach looks good

@dpascualhe