Deep Similarity Optical Flow Estimation Under Lighting Variation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing unsupervised learning methods for optical flow estimation using low-level features like RGB and gradient are not robust to changes in real environments, such as variations in lighting and object angles, leading to inaccurate optical flow estimation.
Innovation Solution
A computer system that utilizes deep similarity by extracting deep features from a sequence of images and learns optical flow through unsupervised learning, employing a feature separation loss function to differentiate occlusion and non-occlusion locations on a deep similarity map, enhancing the performance of optical flow estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If low-level features like RGB and gradient are used for optical flow estimation, then the calculation is simple and fast, but the accuracy and robustness to environmental changes deteriorate
Solution Approach 1:
The patent transforms the feature representation parameters from low-level (RGB, gradient) to high-level deep features extracted through convolutional neural networks. This parameter change enables the system to capture semantic information and maintain robustness under environmental variations while preserving computational efficiency through learned feature hierarchies.
Solution Approach 2:
The patent replaces traditional hand-crafted feature extraction mechanisms with deep learning-based feature extraction. Instead of using fixed mathematical transformations (gradient calculations, RGB comparisons), the system employs trainable convolutional networks that automatically learn optimal feature representations, achieving both accuracy and efficiency.
2Device complexity
If low-level features like RGB and gradient are used for optical flow estimation, then the feature extraction is simple, but the robustness to lighting and angle variations worsens
Solution Approach 1:
The patent changes the feature parameters from static low-level descriptors to dynamic deep features that adapt to environmental conditions. The deep features capture semantic information that remains consistent across lighting and angle variations, significantly improving reliability while the modular network architecture keeps the system manageable.
Solution Approach 2:
The patent combines multiple deep features from different convolutional network layers to create a composite feature representation. This composite approach integrates information from various levels of abstraction, making the optical flow estimation robust to environmental changes while maintaining a structured and manageable system architecture.
3Measurement precision
If deep features are used for optical flow estimation, then the accuracy and robustness improve, but the computational complexity increases
Solution Approach 1:
The patent segments the feature extraction process into multiple convolutional layers, each handling specific aspects of feature detection. This segmentation allows the system to process complex deep features in a modular fashion, improving accuracy through hierarchical feature learning while managing computational complexity through distributed processing across layers.
Solution Approach 2:
The patent introduces dynamic feature extraction through convolutional neural networks that adaptively learn feature representations during training. The dynamic nature of the learned filters allows the system to capture complex patterns efficiently, achieving high accuracy without proportionally increasing computational complexity through parameter sharing and hierarchical processing.
Data Source
AI summary
A computer system of unsupervised learning with deep similarity for optical flow estimation is provided. The computer system is configured to calculate deep similarity by using deep features extracted from a sequence of a plurality of images, and learning optical flow for the images based on the deep similarity.