YOLO Semantic Conditioning — Parking Slot Detection
YOLO26n fine-tuned with CLIP-based semantic conditioning on a parking slot dataset. Text descriptions attached to bounding boxes guide neck features via InfoNCE contrastive loss.
Results
| Model | mAP50 | mAP50-95 |
|---|---|---|
| Baseline (no semantic) | 0.872 | 0.663 |
| This model (semantic_v4) | 0.890 | 0.660 |
+0.018 mAP50 over equal-epoch baseline.
Training Curves
| Baseline | Semantic v4 |
|---|---|
 |
 |
Confusion Matrix
| Baseline | Semantic v4 |
|---|---|
 |
 |
Semantic Alignment Visualization
| Epoch 20 | Epoch 30 | Epoch 40 | Epoch 49 |
|---|---|---|---|
 |
 |
 |
 |
Usage
from ultralytics import YOLO
model = YOLO('best.pt')
results = model('your_image.jpg')
Training Config
tau=0.1— InfoNCE temperaturesem_weight=0.2— semantic loss weightsem_warmup=10— epochs before semantic loss activates- 50 epochs, batch=16, imgsz=640, A100 40GB