YOLOVAL: inspecting saved game detections with sv.Detections (paid ONNX models)

[bbshdh56]

I put together a small Supervision 0.31.4 example for inspecting saved game-image detections from my ONNX models. I'm the creator of YOLOVAL; the model packages are paid, and the overview below is public.

The example takes saved bounding boxes, confidence scores and class IDs and turns them into sv.Detections. It handles empty results, lets me filter by confidence, and keeps full-frame and center-crop detections in the same coordinate system.

For crop results, I first add the crop's left/top offset to each box and save that as xyxy_on_original. Then the conversion is:

import numpy as np
import supervision as sv

def to_detections(rows):
    return sv.Detections(
        xyxy=np.asarray([r['xyxy_on_original'] for r in rows],
                        dtype=np.float32).reshape(-1, 4),
        confidence=np.asarray([r['confidence'] for r in rows], dtype=np.float32),
        class_id=np.asarray([r['class_id'] for r in rows], dtype=np.int32),
    )

Given one view's saved detection records as rows, the inspection steps are:

detections = to_detections(rows)
higher_confidence = detections[detections.confidence >= 0.5]
class_aware = detections.with_nms(threshold=0.45, class_agnostic=False)
centers = detections.get_anchors_coordinates(sv.Position.CENTER)

I ran the converter against existing Delta Force and Arena Breakout ONNX Runtime results: 28 views from 14 source images, containing 25 saved detections and 13 empty views. Empty-array handling and crop-coordinate checks passed. Filtering at 0.5 retained 20 detections; class-aware NMS at 0.45 retained all 25 already-filtered boxes. These are record-processing checks, not an accuracy or speed benchmark. Head and body detections are different classes and can belong to the same character.

Here is the broader 32-example model portfolio, with .onnx filenames and colorful yoloval.cc.cd watermarks. This overview is separate from the 14-image record-processing check.

Open the original 4000 × 2866 image · Model overview repository

Website: https://yoloval.cc.cd/
Patreon shop — paid models, prices and downloads: https://www.patreon.com/c/sugaraigogo/shop