Adaptive Sampling Density for AR Object Detection Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection algorithms for augmented reality (AR) systems require time-consuming manual training processes, involving capturing multiple images of objects from various angles, which is inefficient and limits the immersive experience provided by AR devices like smartphones.
Innovation Solution
A method and system for training object detection algorithms using synthetic two-dimensional (2D) images, where feature data sets from a reference object or 3D model are used to derive similarity scores and vary sampling density based on these scores, generating training data for a head-mounted display (HMD) to detect objects with improved efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual training process is used to capture multiple images of objects from various angles, then object detection accuracy is improved, but training time and effort increase significantly
Solution Approach 1:
The patent uses 3D models as digital copies of physical objects to generate synthetic training images. Instead of manually capturing images of real objects from multiple angles, the system creates virtual representations and renders images from any desired viewpoint, dramatically reducing training preparation time while maintaining detection accuracy
Solution Approach 2:
The patent replaces the mechanical process of physically moving cameras and objects to capture training images with a computational approach. Synthetic images are generated through computer graphics rendering of 3D models, substituting physical image capture mechanics with digital synthesis, thereby eliminating time-consuming manual positioning and capture operations
2Productivity
If uniform sampling density is used across all views, then implementation simplicity is maintained, but training efficiency decreases due to excessive sampling in redundant areas
Solution Approach 1:
The patent implements variable sampling density where different regions of the view sphere have different sampling rates based on their importance. Critical views that provide unique object information are sampled more densely, while redundant views are sampled sparsely, optimizing training efficiency by allocating computational resources locally rather than uniformly
Solution Approach 2:
The patent changes the sampling parameter from a fixed uniform density to a variable density controlled by similarity thresholds. By adjusting the threshold parameter, the system dynamically determines when to insert additional sample views, transforming a static sampling approach into an adaptive one that responds to actual view similarity characteristics
3Measurement precision
If similar adjacent views are sampled densely, then detection accuracy for similar objects is improved, but data redundancy increases
Solution Approach 1:
The patent employs a feedback mechanism where the similarity between adjacent views is continuously evaluated during the sampling process. When similarity exceeds a threshold, the system feedbacks this information to adjust sampling density, stopping further sampling in that region to avoid redundancy while maintaining accuracy where needed
Solution Approach 2:
The patent applies partial sampling action by selectively increasing sampling density only in regions where view similarity indicates potential detection challenges. Rather than uniformly oversampling all views, the system applies excessive sampling locally only where necessary, balancing accuracy requirements with data efficiency
Data Source
AI summary
A head-mounted display, a method, and a non-transitory computer readable medium are provided. An embodiment of a method for obtaining training sample views of an object includes the step of storing, in a memory, multiple views of an object. The method also includes the step of deriving similarity scores between adjacent views and then a sampling density is varied based on the similarity scores.