Face Alignment Model Compression via Cascade Regression Trees
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face alignment approaches are computationally intensive, costly, and power-consuming, and often result in inaccurate predictions due to global training without considering variations in facial pose, lighting, expression, and occlusion, leading to over-fitting issues.
Innovation Solution
A model using a cascade of regression trees with ordered, sequential levels, where each level consists of decision trees that evaluate and adjust the facial shape based on pixel values, utilizing barycentric coordinates for efficient pixel location determination and adjusting learning factors to prevent over-fitting, and compressing the model into a dictionary of atoms for reduced storage and transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a cascade of gradient boosted decision trees or Gaussian process regression trees is used for face alignment, then facial anchor points can be localized, but computational cost and power consumption increase significantly
Solution Approach 1:
The prediction model is divided into multiple cascade levels, where each level contains multiple decision trees that process the input image sequentially. This segmentation allows the system to stop processing early when confidence is high, reducing overall computational cost and power consumption while maintaining localization accuracy.
Solution Approach 2:
The system performs partial processing by evaluating only a subset of decision trees at each cascade level before determining whether to proceed to the next level. This partial action approach reduces the total number of computations required compared to evaluating all trees at all levels, thereby lowering power consumption.
2Measurement precision
If transformation matrix calculation is performed between current coordinates and default shape coordinates for each facial anchor point, then accurate positioning is achieved, but computational intensity increases
Solution Approach 1:
The prediction model is pre-trained offline to learn the mapping between image coordinates and default shape coordinates. During runtime, the system directly applies the learned predictions without performing intensive transformation matrix calculations, thereby maintaining positioning accuracy while significantly reducing computational power requirements.
3Device complexity
If prediction models are trained globally without considering variations in facial pose, lighting, expression, and occlusion, then training complexity is reduced, but prediction accuracy deteriorates
Solution Approach 1:
The training process incorporates local variations by using diverse training data that includes different facial poses, lighting conditions, expressions, and occlusions. The model learns to adapt to these local variations while maintaining a unified global structure, thereby improving prediction accuracy without excessively increasing training complexity.
4Reliability
If a global learning factor is applied to all levels of the cascade to reduce over-fitting, then over-fitting is reduced, but shape prediction accuracy decreases
Solution Approach 1:
The learning factor is made dynamic and level-dependent, with different learning rates applied at different cascade levels. Earlier levels use larger learning factors to learn coarse features quickly, while later levels use smaller learning factors to fine-tune predictions and reduce over-fitting. This dynamic approach maintains shape prediction accuracy while preventing over-fitting.
Data Source
AI summary
A face tracking system generates a model for extracting a set of facial anchor points on a face within a portion of a face image based a multiple-level cascade of decision trees. The face tracking system identifies a mesh shape adjusted to an image of a face. For each decision tree, the face tracking system identifies an adjustment vector for the mesh shape relative to the image of the face. For each cascade level, the face tracking system combines the identified adjustment for each decision tree to determine a combined adjustment vector for the cascade level. The face tracking system modifies adjustment of the mesh shape to the face in the image based on the combined adjustment vector. The face tracking system reduces the model to a dictionary and atom weights using a learned dictionary. The model may be more easily transmitted to devices and stored on devices.