Estimator Training Using Synthetic Depth Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for estimating the pose of an object from depth images are hindered by noise and holes, leading to incorrect pose estimation, particularly when the object is tilted or has varying viewpoints.
Innovation Solution
A method is developed to train an estimator using a combination of synthetic and real data, associating labeled synthetic data with real data to fill in errors and noise, and employing a latent regression forest to accurately estimate the pose of articulated objects like hands, even with incomplete or noisy input images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional methods use depth images for pose estimation, then the process is simple, but holes and noise in the depth images cause incorrect pose estimation
Solution Approach 1:
The patent creates a synthetic depth image that copies the structure and characteristics of the original depth image but without holes and noise. This synthetic copy is then used to train an estimator that can handle real depth images with defects, effectively transferring the perfect structure from the synthetic copy to improve estimation accuracy on imperfect real data.
Solution Approach 2:
The patent performs preliminary training of an estimator using synthetic depth images before deploying it on real depth images. This preliminary action prepares the estimator to handle holes and noise by learning from complete, noise-free synthetic data, enabling it to compensate for information loss in real images during actual pose estimation.
2Measurement precision
If more real data is collected for training, then training accuracy may improve, but resource requirements increase
Solution Approach 1:
Instead of collecting large volumes of real training data, the patent creates synthetic copies of depth images with known poses and structures. These synthetic copies serve as training data, replacing the need for extensive real data collection while maintaining training effectiveness, thus reducing resource requirements.
Solution Approach 2:
The patent changes the parameters of the training data by using synthetically generated depth images with controlled characteristics rather than raw real-world data. This parameter change allows for efficient training with fewer resources while achieving comparable or superior accuracy through the controlled perfection of synthetic data.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A method and an apparatus for training an estimator for estimating a pose of an articulated object are disclosed. The method of training an estimator comprises: obtaining real object data from an input image of an object; independently generating synthetic object data using a model of the object; associating the synthetic object data with the real object data; and training, by way of a processor, the estimator to estimate a pose of the object based on the associating of the synthetic object data and the real object data. The input image of the object may comprises a depth image.