Face 98-Keypoint Recognition
Face 98-keypoint recognition is an AI algorithm that detects 98 facial feature points with high precision from face images, including eyebrows, eyes, nose, mouth, and facial contour. It serves as a foundational technology for face-related applications such as high-precision face recognition, facial beautification and editing effects, expression analysis, and lip reading.
Algorithm Overview
It consists of two stages: face detection and keypoint estimation.
- Face Detection: Identifies the face region from the image
- 98-Keypoint Detection: Detects a total of 98 points including eyebrows (x10), eyes (x10), nose (x9), mouth (x20), facial contour (x33), and pupils (x2)
Performance Metrics
| Dataset | Error (NME %) |
|---|---|
| 300W | 2.78 |
| COFW | 3.08 |
| AFLW | 1.42 |
Edge AI Board (RV1126B) Execution Efficiency
| Processing Stage | Model Size | Processing Time |
|---|---|---|
| Face Detection (face_detect) | 44.23MB | 17ms |
| 98-Keypoint Detection (face_landmark98) | 10.88MB | 23ms |
| Total | 55.11MB | Approx. 40ms |
Key Features
- High-density keypoints: High-precision detection of 98 points across the entire face
- Multi-benchmark validation: 300W (NME 2.78), COFW (NME 3.08), AFLW (NME 1.42)
- Real-time performance: Detection + keypoints combined in approximately 40ms
- Diverse applications: Face recognition, beautification, expression analysis, gaze estimation, 3D face reconstruction
Use Cases
- High-precision face recognition systems (accuracy improvement through facial pose correction)
- Beauty and cosmetic apps (per-feature makeup and correction)
- Expression analysis and emotion recognition
- AR face filters (social media effects)
- 3D face model reconstruction
- Lip reading (lip-reading, speech detection)
- Driver drowsiness and distraction detection
Edge AI Board Implementation
Using the RV1126B NPU, face detection (17ms) and 98-keypoint detection (23ms) are processed in a combined total of approximately 40ms. Performance is sufficient to keep up with real-time video streams.