A face mask detection system built with Faster R-CNN to classify three categories — proper mask wearing, no mask, and incorrectly worn mask. This was built as a learning project to understand two-stage object detection on a small real-world dataset.
| Metric | Value |
|---|---|
| mAP@0.5 | 86.5% |
| Training Epochs | 15 (early stopped) |
| Best Val Loss | 0.2803 |
| Final Train Loss | 0.0959 |
| Dataset Size | 853 images |
| Train/Val/Test Split | 597/127/129 (70/15/15) |
Source: Kaggle Face Mask Detection Dataset
- Total Images: 853
- Total Bounding Boxes: 4,072
- Format: PASCAL VOC XML annotations
- Classes: 3 (with_mask, without_mask, mask_weared_incorrect)
with_mask: 3,232 (79.37%)
without_mask: 717 (17.61%)
mask_weared_incorrect: 123 ( 3.02%)
Note: The dataset has a natural class imbalance. The minority class mask_weared_incorrect (3%) is the weakest performer in the model. We relied on COCO-pretrained transfer learning to partially handle this, but collecting more examples of this class is the proper long-term fix.
| Decision | Reasoning |
|---|---|
| Two-stage detector | Chosen specifically to learn and understand how two-stage object detection works — RPN + classification head |
| ResNet50-FPN backbone | FPN handles objects at multiple scales well. However in hindsight, ResNet18-FPN would have been a better fit for a dataset this small — less complexity, less overfitting risk |
| Transfer learning from COCO | Pretrained weights reduced training time and helped the model generalize despite the small dataset |
| Proven architecture | Well documented, easier to debug and understand compared to newer methods |
- Batch Size: 4 (limited by GPU memory — 6GB VRAM)
- Learning Rate: 5e-4 with ReduceLROnPlateau scheduler
- Optimizer: Adam
- Early Stopping: Patience of 7 epochs
- Random Seed: 42 (for reproducibility)
RandomHorizontalFlip(0.5) # Faces can appear mirrored
ColorJitter(brightness=0.2, # Handle different lighting conditions
contrast=0.2,
saturation=0.1)Augmentation was important here because 853 images is a very small dataset. These transforms increase effective variety without collecting new data.
Batch Size 4:
- Tested 2, 4, and 8
- Size 8 caused GPU out-of-memory errors
- Size 2 caused unstable gradients
- 4 was the stable middle ground
Learning Rate 5e-4:
- 1e-3 — loss diverged (jumped around, never settled)
- 5e-4 — stable convergence
- 1e-4 — converged but very slowly
- Scheduler automatically reduced to 2.5e-4 at epoch 12
70/15/15 Split:
- Separate test set ensures unbiased final evaluation
- Fixed seed guarantees same split every run
- No data leakage between sets
Epochs 1–8: Both train and val loss decreasing steadily Epoch 8: Best validation loss achieved (0.2803) Epochs 9–11: Val loss starts rising while train loss keeps dropping → overfitting begins Epoch 12: Scheduler triggers, LR reduced from 5e-4 to 2.5e-4 Epochs 12–15: No improvement even with lower LR Epoch 15: Early stopping triggered
Train Loss: 0.0959 ← very low
Val Loss: 0.3171 ← 3.3x higher
The model memorized the training data rather than learning features that generalize.
- Dataset too small — 853 images is not enough variety for a model this size. Object detectors typically need 5,000+ images
- Model too complex — ResNet50 has too many parameters for this dataset. ResNet18 would have been a better backbone choice
- Limited data diversity — mostly frontal faces, similar lighting, similar contexts
- Unusual angles — trained mostly on frontal faces, struggles with side profiles and tilted heads
- False positives — sometimes detects masks on ears, hair, or background objects
- Occlusions — hands covering face or sunglasses combined with mask confuse the model
- Minority class —
mask_weared_incorrecthas very few examples, model underperforms on it
- Replace ResNet50 with ResNet18 backbone — reduces model complexity and overfitting risk
- Add dropout layers for regularization
- More aggressive augmentation (rotations, perspective changes, synthetic occlusions)
- Collect 5,000+ diverse images with varied angles, lighting, and demographics
- Add focal loss to better handle the minority class imbalance
- Once dataset is larger, explore more powerful architectures
git clone https://github.com/Pooja-Vachhad/face-mask-detection.git
cd face-mask-detection
pip install -r requirements.txtpython train.py- Max 30 epochs with early stopping
- Best model saved to
best_model.pth - ~1.5 hours on T4 GPU
python test.py- Model size must match dataset size — ResNet50 was overkill for 853 images
- Transfer learning helps a lot on small datasets — COCO pretraining gave a strong starting point
- Early stopping is necessary but not sufficient — it limits overfitting but doesn't fix the root cause
- Train loss alone is misleading — always watch validation loss
- Start with the simplest model that makes sense — don't go heavy by default
- Document what went wrong honestly — it shows deeper understanding than hiding problems
- mAP on test set ≠ production ready — edge cases matter in real deployment
YOLO is significantly faster but this project was specifically about learning how two-stage detection works. Understanding the RPN, anchor boxes, and ROI pooling in Faster R-CNN was the goal not achieving maximum inference speed.
Binary mask/no-mask misses a real common case mask worn with nose exposed. Three classes gives more actionable output and reflects what you actually see in practice.
We used COCO-pretrained transfer learning as the primary strategy for handling imbalance. The minority class mask_weared_incorrect still underperforms this is acknowledged as a known limitation. The real fix is more data, not reweighting.