This repository provides a practical evaluation and comparison of popular face detection models under various real-world conditions. The tested models include Haar Cascade, dlib (HOG+SVM and CNN), MTCNN, YOLO variants (v11s and v11n), and RetinaFace.
The accuracy and performance of each model were evaluated across diverse scenarios:
- Angles: Different facial orientations and positions.
- Lighting Conditions: Varied lighting intensities, including bright, medium, and low-light environments.
- Distances: Short, medium, and long-range detections.
- Accessories: Impact of wearing sunglasses and masks on face detection accuracy.
- Multiple Faces: Efficiency in detecting more than one face simultaneously.
| Model | Speed | Detection Distance | Angle Stability | Bright Light | Low Light | Mask | Glasses | Mask & Glasses | Multiple Faces | Detection Accuracy |
|---|---|---|---|---|---|---|---|---|---|---|
| Haar Cascade | > 30 fps | ≤ 2 meters | Very Weak | Moderate | Weak | Weak | Good | Very Weak | Weak | Weak |
| dlib-HOG+SVM | > 30 fps | ≤ 1 meter | Weak | Good | Weak | Weak | Good | Very Weak | Weak | Weak |
| dlib-CNN | < 1 fps | - | - | - | - | - | - | - | Very Good | Very Good |
| MTCNN | 5-7 fps | > 3 meters | Very Good | Very Good | Very Good | Very Good | Very Good | Moderate | Very Good | Very Good |
| Yolo v11s | 10-12 fps | > 3 meters | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good |
| Yolo v11n | 15-17 fps | > 3 meters | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good |
| RetinaFace | 1 fps | > 3 meters | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good | Very Good |
all speed measurements were performed on CPU only, using the following system: 11th Gen Intel(R) Core(TM) i5-11400H @ 2.70GHz
- Real-time with multiple people: YOLO v11n
- Offline high-accuracy analysis: RetinaFace
- Low-resource, fast preview: Haar Cascade (for basic needs only)
- Balanced accuracy & speed: MTCNN / YOLO v11s
This resource is particularly valuable for:
- Researchers and students evaluating face detection methods.
- Developers and engineers selecting appropriate models for real-world face detection applications.
- Open-source contributors aiming to enhance model robustness.
This repository is licensed under the MIT License. Feel free to explore, contribute, or utilize the insights provided here to inform your face detection projects.