Depth Image Similarity Checking for Regression Pose Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information processing systems struggle to effectively specify erroneously recognized images in regression tasks, making it difficult to automatically execute processing based on recognition results.
Innovation Solution
An information processing apparatus and method that utilizes depth images, pose estimation, and 3D computer graphics to calculate similarity between actual and rendered images, enabling the identification and correction of erroneous recognition through a combination of similarity calculation and classification techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If majority decision is used to specify erroneously recognized images in classification tasks, then recognition accuracy can be improved, but the method cannot be directly applied to regression tasks
Solution Approach 1:
The patent changes the parameter used for evaluation from classification-based majority decision to regression-based similarity measurement. By using similarity calculators that compute similarity scores between predicted and actual values, the system adapts the error specification method to work with regression tasks where output values are continuous rather than categorical.
Solution Approach 2:
The patent replaces the mechanical majority decision voting system with a similarity-based evaluation mechanism. Instead of counting votes from multiple classifiers, the system uses similarity calculators to measure the degree of match between predicted regression outputs and actual values, substituting a quantitative similarity metric for the qualitative majority decision process.
2Productivity
If automatic collection of erroneously recognized images is implemented, then learning efficiency is improved, but system complexity increases
Solution Approach 1:
The patent implements self-service by enabling the system to automatically identify and collect its own errors without external intervention. The similarity calculators continuously monitor performance, automatically flag erroneously recognized images based on similarity thresholds, and facilitate automatic collection for relearning, making the system self-diagnostic and self-improving.
Solution Approach 2:
The patent establishes a feedback loop where similarity measurements provide continuous information about recognition accuracy. This feedback mechanism allows the system to identify errors in real-time, collect erroneous images for analysis, and use these insights to improve future recognition performance, creating a closed-loop learning system.
Data Source
AI summary
Provided is an information processing apparatus capable of automatically executing processing according to a recognition result from a regression task. An acquirer acquires a first depth image that is a depth image of an area including a subject. Next, a pose estimator estimates coordinates of a plurality of feature points predetermined for the subject on the basis of the acquired first depth image. Then, a rendering unit generates a 3D CG model of the subject on the basis of the coordinates of the estimated feature points, and renders a second depth image obtained by imaging the generated 3D CG model with a virtual camera. Thereafter, a similarity calculator calculates similarity between the acquired first depth image and the rendered second depth image. Thereafter, the processing executor executes processing according to the calculated similarity.