Circuit Component Detector (YOLO26M-OBB)
YOLO26M-OBB model trained for detecting electronic schematic components in hand-drawn circuit diagrams.
Performance
| Metric |
Value |
| mAP50 |
88.5% |
| mAP50-95 |
78.3% |
| Precision |
95.6% |
| Recall |
88.6% |
Per-Class Recall
| Class |
Recall |
mAP50 |
| operational_amplifier |
100.0% |
99.4% |
| inductor |
94.9% |
94.7% |
| voltage_source |
92.8% |
95.3% |
| capacitor |
92.2% |
92.7% |
| transistor |
92.1% |
94.8% |
| resistor |
91.7% |
93.2% |
| diode |
91.6% |
93.1% |
| integrated_circuit |
91.3% |
91.0% |
| other |
90.4% |
93.4% |
| gnd |
89.4% |
88.9% |
| text |
85.3% |
85.2% |
| junction |
84.9% |
84.3% |
| terminal |
84.2% |
84.9% |
| switch |
83.4% |
78.3% |
| vss |
83.1% |
85.9% |
| crossover |
70.7% |
69.3% |
Usage
Classes (16)
| ID |
Class |
| 0 |
resistor |
| 1 |
capacitor |
| 2 |
diode |
| 3 |
transistor |
| 4 |
inductor |
| 5 |
voltage_source |
| 6 |
integrated_circuit |
| 7 |
operational_amplifier |
| 8 |
other |
| 9 |
gnd |
| 10 |
text |
| 11 |
junction |
| 12 |
terminal |
| 13 |
switch |
| 14 |
vss |
| 15 |
crossover |
Training Details
- Model: YOLO26M-OBB
- Dataset: CGHD-1152 (61 classes → 16 merged)
- Split: 85/15 random (2,652 train / 468 val)
- Excluded: drafter_0 (different drawing style)
- Epochs: 200
- Augmentations: mosaic=1.0, mixup=0.15, degrees=10, translate=0.2, scale=0.5, shear=2, fliplr=0.5, flipud=0.1, erasing=0.4, hsv, randaugment
- Optimizer: AdamW, lr0=0.001, cos_lr=true
- Image size: 1024
Key Learnings
- Class merging critical: 61→16 classes improved mAP from ~50% to 85%
- Augmentations help: +3.5% mAP over no-augmentation baseline
- M model > L model: Smaller model generalizes better on this dataset size
SHA256
