YoloPiece One-Shot Models

Unified browser package for Poneglyph one-shot analysis.

This bundle replaces the previous ReaderNet runtime with two small pairwise rankers. The bubble detector is preserved, and the panel detector is included because the worker assigns bubbles to panels before in-panel ordering.

Files

• bubble_detector.onnx: preserved YOLO bubble detector.
• panel_detector.onnx: YOLO panel detector used to assign bubbles to panels.
• panel_order.onnx: new pairwise panel ordering ranker.
• bubble_order.onnx: new pairwise in-panel bubble ordering ranker.
• global_bubble_order.onnx: optional page-level reranker, present only when benchmark-enabled.
• model_manifest.json: checksums, sizes, and runtime metadata.
• metrics/: held-out evaluation reports.

Reading Order Metrics

• Test pages: 31
• Panel exact order: 1.0000
• Bubble exact order inside panels: 0.9596
• Page full accuracy: 0.8387 (26/31)

The page full-accuracy metric compares the complete predicted bubble sequence against the ground truth sequence for held-out pages.

Detector Metrics

• Panel detector mAP50: 0.9940
• Panel detector mAP50-95: 0.9861

Runtime

The frontend worker loads these files from:

https://huggingface.co/Remidesbois/YoloPiece_OneShot_Models/resolve/main/<file>

Optional Global Bubble Reranker

• global_bubble_order.onnx: enabled by the completed benchmark.
• Feature count: 108
• Selected experiment: global_bubble_reranker_v1__sklearn-logistic__borda-vertical-repair
• Page full accuracy: 0.9677
• Global pairwise accuracy: 0.9993