Spatial Relationship Labeling With Reinforcement Learning Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision techniques for estimating spatial relationships between images, such as optical flow, disparity, and depth, are computationally inefficient due to the use of exhaustive or randomized search methods, leading to high computational costs.
Innovation Solution
A reinforcement learning model is employed to predict pixelwise spatial relationships between images, utilizing attention mechanisms tuned to different time scales and optimizing labeling maps through reinforcement learning techniques, improving the efficiency of correspondence estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exhaustive or randomized search methods are employed to estimate pixel correspondence, then measurement precision is improved, but productivity deteriorates due to significant compute time and resources consumption
Solution Approach 1:
The patent segments the pixel correspondence estimation problem into discrete spatial relationship labels (e.g., left, right, above, below, etc.) that can be independently predicted. Instead of searching through all possible pixel correspondences, the system divides the continuous correspondence space into discrete directional categories, enabling efficient classification-based prediction while maintaining acceptable precision.
Solution Approach 2:
The patent transforms the pixel correspondence estimation from a continuous coordinate matching problem into a discrete classification problem with predefined spatial relationship labels. By changing the parameter space from continuous pixel coordinates to discrete directional labels, the system achieves both computational efficiency and measurement precision through learned probability distributions.
2Measurement precision
If reinforcement learning with iterative shift is used for visual tracking, then measurement precision of spatial relationships is improved, but device complexity increases due to multiple updating cycles
Solution Approach 1:
The patent implements dynamic iterative updating of the reinforcement learning model where the spatial relationship labeling is refined through multiple cycles. Each iteration shifts the labeling based on updated predictions, allowing the model to adapt and improve precision dynamically. The system balances this increased complexity by using efficient probability distribution updates and learned features from the neural network.
Data Source
Figure 1
Figure 2~4
Figure 5
AI summary
A computing system is provided, including a processor configured to receive a labeling map for a first image. The labeling map may indicate a spatial relationship between a first region of interest included in the first image and a second region of interest included in a second image. At a trained reinforcement learning model, the processor may be further configured to generate an updated labeling map for the first image based on at least the labeling map, the first image, and the second image.