Face Bounding Box Estimation From Keypoints on Low-Resource Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional face detection techniques using neural networks are resource-intensive, making them challenging to implement on low-budget or size-constrained devices, and are not suitable for accurately detecting animal faces due to differences in facial topologies and features.
Innovation Solution
An electronic device uses facial keypoints to detect and estimate the pose of animal faces, computing the size and position of a bounding box through computational geometry, reducing computational and memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural network models are used for face detection, then detection accuracy is improved, but computational resource consumption and memory requirements increase significantly
Solution Approach 1:
The patent segments the face detection task into two distinct phases: (1) a lightweight initial detection phase using simple algorithms to identify potential face regions, and (2) a refinement phase that applies more sophisticated analysis only to these candidate regions. This segmentation reduces overall computational load while maintaining detection accuracy, as the complex processing is applied only where needed rather than to the entire image.
Solution Approach 2:
The patent extracts and utilizes specific geometric features (triangulation relationships between facial landmarks, contour characteristics) from face images to create a dedicated feature representation. By extracting these key geometric properties and using them as the basis for detection and recognition, the system achieves accurate face detection without requiring full neural network processing of the entire image, thereby reducing computational resource consumption.
2Adaptability or versatility
If neural network models are deployed on low-budget or size-constrained devices, then face detection functionality is achieved, but other parallel operations are constrained due to memory and computational resource consumption
Solution Approach 1:
The patent applies local quality by using a lightweight detection algorithm for the initial face region identification that consumes minimal computational resources, and then applying more resource-intensive geometric feature analysis only locally within the detected candidate regions. This localized application of sophisticated processing preserves memory and computational resources for other parallel operations on the device while still achieving accurate face detection functionality.
3Measurement precision
If conventional face detection techniques are used, then human face detection is achieved, but animal face detection accuracy is insufficient due to differences in facial topologies and features
Solution Approach 1:
The patent creates a universal geometric feature extraction framework based on triangulation relationships between facial landmarks that can be applied to both human and animal faces. By focusing on fundamental geometric properties (relative positions, distances, and angular relationships between key facial points) rather than species-specific features, the system achieves adaptability across different facial topologies while maintaining detection accuracy.
Solution Approach 2:
The patent adapts the detection parameters and geometric feature weights based on the detected species type. For example, the relative importance of different facial landmarks, the scaling factors for triangulation calculations, and the threshold values for recognition are adjusted according to whether the detected face is human or animal. This parameter adaptation allows the same core algorithm to achieve high accuracy across different species despite their facial topology differences.
Data Source
AI summary
An electronic device and method for keypoints-based estimation of face bounding box is provided. The electronic device receives an image that includes an object and detects a set of facial keypoints on a face of the object in the image. The set of facial keypoints corresponds to a set of facial features of the object. The electronic device identifies a face pose of the object in the image based on the detected set of facial keypoints and computes a set of values that is associated with a size and a position of a face bounding box in the image. The computation is based on the identified face pose and locations of the set of facial keypoints in the image. The electronic device overlays the face bounding box on the image based on the computed set of values.