Stereo Optical Flow Partitioning for High-Resolution Memory Limits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based optical flow estimation for high-resolution images is constrained by memory limitations, leading to undesirable artifacts when down sampling and up sampling is employed, which degrades the quality of view synthesis results.
Innovation Solution
A multi-level optical flow framework that partitions input images into regions suitable for deep learning processing at full resolution, merging the results to generate high-quality optical flow maps without downsampling, using techniques like convolutional neural networks and filtering to smooth seams.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning-based optical flow estimation is applied to high-resolution images, then estimation accuracy is improved, but memory constraints prevent direct processing
Solution Approach 1:
The patent divides the high-resolution image into multiple lower-resolution patches or tiles that can be processed independently by the deep learning model. Each patch is processed separately to generate optical flow estimates, which are then merged to produce the final high-resolution optical flow map. This segmentation allows the memory-constrained model to handle high-resolution images without requiring excessive memory capacity.
2Ease of manufacture
If images are downsampled to meet memory constraints, then processing becomes feasible, but artifacts are introduced in the output
Solution Approach 1:
By processing image patches at their native resolution rather than downsampled resolution, the patent avoids the artifacts introduced by downsampling. Each patch is processed independently at full resolution, and the results are merged to produce a high-quality optical flow map without the degradation associated with global downsampling and upsampling operations.
Solution Approach 2:
The patent transitions from processing the entire image at once to processing multiple smaller patches in parallel. This dimensional change from single large-image processing to multi-patch processing allows the system to maintain high resolution while meeting memory constraints, eliminating the need for quality-degrading downsampling.
3Measurement precision
If entire high-resolution images are processed at once, then full resolution results are achieved, but memory constraints are exceeded
Solution Approach 1:
The patent segments the high-resolution image into multiple smaller patches that can be processed in parallel by the deep learning model. Each patch is processed at its native resolution to maintain quality, and the optical flow estimates from all patches are merged to produce the final high-resolution result. This segmentation strategy achieves full resolution output while keeping individual memory requirements manageable.
Data Source
AI summary
Techniques related to multi-level optical flow estimation are discussed. Such techniques include partitioning each pair of input images into one or more partitions, separately performing optical flow estimation on the partitions, and merging the separately generated optical flow results into a final optical flow map for the pair of input images.