Multimodal Technical Drawing Extraction

v1.0 Frozen

Zero-Downscale YOLOv8 Layout Decomposition + Chandra 2 INT4 Multimodal OCR

Sheet:
Strictness:
Chandra CER
2.3% Levenshtein
Text EMR
91.4% Zero defect
Dimension Match
93.5% 820/877
BOM Cell Match
92.9% 19.1k/20.6k
YOLO mAP@0.5
84.9% 50-95: 74.4%
Precision / Recall
86.5% / 84.6%
161 FP • 19.1ms
IoU Multi-Tier Gating
≥0.90 62.9%
≥0.75 82.8%
≥0.50 91.8%
Benchmark Corpus
80 Sheets
477 Crops Verified
100% | [0, 0] px
#01 Loading Sheet...
Original Annotations (Ground Truth)
Master Drawing
Table (#10b981) Title Block (#3b82f6) Diagram (#f97316) Notes (#8b5cf6)