bev-ens1
Beverage detection for Bittensor subnet 44 (Score), public track.
Two YOLO26n detectors are run on one shared 704x704 letterboxed tensor, reduced to their own detections by per-class NMS, then combined by weighted box fusion. The two models disagree on different images, and fusing them scores higher than either alone.
Contents
| file | what it is |
|---|---|
weights.onnx |
derived from WhiteCrow323/beverage1 @f33d89f5 |
weights_b.onnx |
derived from Realfencer/bev918-1 @37ff9b53 |
miner.py |
the ensemble serving path |
chute_config.yml |
deployment spec |
Both source models are published openly on the subnet and are Ultralytics
YOLO26n exports under AGPL-3.0; this derivative keeps that licence. Classes are
the element's own order: cup, can, bottle.
Configuration
Per-model NMS at IoU 0.45, fusion at IoU 0.45, fused confidence 0.30. Those were chosen by five-fold cross-validation on 243 challenge images, with thresholds picked on four folds and read off the fifth.
Inference threads are min(8, os.cpu_count()) rather than a fixed 2, because the
audit sandbox sets no CPU limit.