Using box_non_max_merge to Merge YOLO Boxes?

[DP1701]

Search before asking

• I have searched the Supervision issues and found no similar feature requests.

Question

Hello everyone,

maybe someone can help me out. I’m facing the following challenge: In my scene, a truck drives into an area and moves toward an unloading zone (always exactly one truck). However, YOLO returns multiple detection boxes for the same truck — for example parts of the trailer, the cab, or parts of the cargo. In principle the detection works, but I’d like to merge these boxes into a single one since they all overlap anyway.

In Supervision, I found the function box_non_max_merge(), which I think could solve this problem. However, the function doesn’t expect a Detections object but rather a 2D array of object detection predictions.

My current approach looks like this:

import supervision as sv
from supervision.detection.core import merge_inner_detections_objects_without_iou

preds = np.hstack((detections.xyxy, detections.confidence.reshape(-1, 1), detections.class_id.reshape(-1, 1) ))

merge_groups = sv.box_non_max_merge( predictions=preds, iou_threshold=0.2, overlap_metric=sv.OverlapMetric.IOU)

merged_detections = []
for group in merge_groups:
      per_detection = [detections[i] for i in group]
      merged_detections.append(merge_inner_detections_objects_without_iou(per_detection))

 detections = (
       sv.Detections.merge(merged_detections)
       if merged_detections
       else sv.Detections.empty()
   )

Is this the intended way to use the function, or are there better methods to merge all overlapping boxes into a single truck bounding box?

Additional

No response

[DP1701]

Okay, I should have searched the Doc more thoroughly. Everything is easier with:

detections = detections.with_nmm(threshold=0.2, overlap_metric=sv.OverlapMetric.IOU)

If anyone ever faces the same challenge.

[Borda]

Using detections.with_nmm() is the intended and most efficient way to handle this in supervision.

Your initial approach involved manually calling the low-level utility box_non_max_merge, which requires unpacking the Detections object. The with_nmm method acts as a wrapper around that utility, handling the unpacking and repacking of the Detections object for you while preserving tracker IDs, class IDs, and other attributes automatically.

This feature was highlighted in the v0.21.0 release—you can read more about it in the release notes here.