Optical Flow Estimation via Low-Resolution Inference and Guided Upsampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural network-based optical flow estimation methods are limited for real-time computer vision applications due to high computational requirements and time consumption.
Innovation Solution
A method involving a two-stage approach: first estimating optical flow at a lower resolution using a neural network and then upsampling it to the original resolution using a guided neural network, leveraging downscaling and adaptive upsampling techniques to maintain accuracy while reducing computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are used for optical flow estimation, then accuracy is improved, but computational resources and time consumption increase significantly
Solution Approach 1:
The patent divides the optical flow estimation task into two separate stages: first estimating optical flow at a lower resolution to capture coarse motion patterns, then refining the result at full resolution. This segmentation allows the system to benefit from both low-resolution speed and high-resolution accuracy, resolving the contradiction between processing speed and estimation accuracy.
Solution Approach 2:
The patent introduces a resolution dimension by performing optical flow estimation at multiple resolutions (low resolution first, then full resolution). This dimensional approach allows the system to process information at different levels of detail, achieving both speed (through low-resolution processing) and accuracy (through full-resolution refinement).
2Measurement precision
If deep neural networks process full-resolution image frames, then optical flow accuracy is maintained, but computational load increases
Solution Approach 1:
The patent applies partial action by performing the computationally intensive deep neural network processing only at low resolution, rather than at full resolution. This partial processing approach captures the essential motion information while significantly reducing computational resource consumption and energy usage.
Solution Approach 2:
The patent extracts and processes only the essential motion information at low resolution, separating this from the full-resolution image data. By taking out the motion estimation task and handling it at reduced resolution, the system minimizes computational resource usage while preserving the critical optical flow information.
Data Source
AI summary
Systems and methods are provided for optical flow estimation. In one embodiment, a method comprises estimating, with a first neural network, an optical flow between two image frames, wherein a resolution of the optical flow is lower than a resolution of the two image frames, and upsampling, with a second neural network, the optical flow to the resolution of the two image frames. In this way, the speed of optical flow estimation may be improved by reducing the amount of pixels being processed by a deep neural network, while the use of another deep neural network for guided upsampling of the optical flow estimate helps maintain the accuracy of the final output.