After training prediction give double bounding box

Support
, ,
1

Good morning,

I’m trying to train a yolo v11n model for defect detection in images coming from manifacturing application. The problem has only one class to detect.

After the training I have a strange behaviour. If I test the best produced model on the dataset, I always have back from the model an almost perfect detection (compared to ground truth) and a second one that is always next to the correct one. For example, the correct bounding box is [122, 246, 11, 10] and thr fake one is [133, 246, 11, 10]. The strange thing is that the score is very similar between the two output boxes, and also the class is always the same (0 in my case since I have only one class). The behaviour is the same on all the image taken from train, validation and test set.

I checked all the label file and everything is correct. Any idea about the possible issue?

Regards

2

Please share your inference code that’s being used. Without the code it will be challenging to help diagnose the issue.

3

I use the code from tutorial

model = YOLO(file.pt)

results = model([“img.jpg“], conf=0.3)

for result in results:

    boxes = result.boxes
    sprint(boxes)

This is the typical output

0: 640x640 2 defects, 9.1ms
Speed: 1.0ms preprocess, 9.1ms inference, 1.2ms postprocess per image at shape (1, 3, 640, 640)
ultralytics.engine.results.Boxes object with attributes:

cls: tensor([0., 0.], device=‘cuda:0’)
conf: tensor([0.4887, 0.4078], device=‘cuda:0’)
data: tensor([[116.5595, 240.8561, 128.1036, 251.5495, 0.4887, 0.0000],
[128.0958, 240.5381, 139.5009, 251.5142, 0.4078, 0.0000]], device=‘cuda:0’)
id: None
is_track: False
orig_shape: (640, 640)
shape: torch.Size([2, 6])
xywh: tensor([[122.3316, 246.2028, 11.5441, 10.6934],
[133.7984, 246.0262, 11.4050, 10.9761]], device=‘cuda:0’)
xywhn: tensor([[0.1911, 0.3847, 0.0180, 0.0167],
[0.2091, 0.3844, 0.0178, 0.0172]], device=‘cuda:0’)
xyxy: tensor([[116.5595, 240.8561, 128.1036, 251.5495],
[128.0958, 240.5381, 139.5009, 251.5142]], device=‘cuda:0’)
xyxyn: tensor([[0.1821, 0.3763, 0.2002, 0.3930],
[0.2001, 0.3758, 0.2180, 0.3930]], device=‘cuda:0’)

4

Thanks for the details and the printout. What you’re seeing is expected when two small, side-by-side boxes don’t overlap enough for NMS to suppress one of them. NMS filters by IoU, so if IoU≈0, both survive. Your two boxes are almost touching in x with tiny/no overlap, so standard NMS won’t remove the “twin” box. A quick primer is in our article on Non‑Maximum Suppression (NMS) explained.

Quick options:

  • If there is at most one defect per image, keep only one: results = model('img.jpg', conf=0.3, iou=0.7, max_det=1).
  • If you need multiple defects per image, post-process by merging near-duplicate boxes by center proximity (not IoU), keeping the highest‑conf in each cluster. You can replace boxes in-place with result.update(boxes=new_boxes) as shown in the Results API docs.
  • Improving localization often removes these splits: try a larger imgsz (e.g., imgsz=1024) or a slightly larger model size (YOLO11s) and verify on the latest ultralytics release.

If you can share one example image + its label and your ultralytics.__version__, I can try to reproduce and suggest exact thresholds for the proximity merge.