Synthetic Frame Training for Difficult Object Motion Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current state-of-the-art DNN-based motion estimation methods struggle with challenging cases where objects undergo severe appearance variation or small, fast-moving objects, as they often fail to accurately estimate motion across frames.
Innovation Solution
The method involves segmenting frames to identify potential object-instances, extracting feature vectors, and matching these objects between frames using clustering and object confidence scores. Additionally, a training dataset is generated to improve the accuracy of motion estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If DNN-based methods are used for motion estimation, then computational efficiency is improved, but accuracy deteriorates in challenging cases with severe appearance variation or small fast-moving objects
Solution Approach 1:
The patent segments the image into multiple regions and processes each region separately through the DNN. This allows the system to apply different processing strategies to different parts of the image, improving accuracy for challenging regions (like small fast-moving objects) while maintaining computational efficiency for easier regions. The segmentation enables focused attention on difficult areas without processing the entire image at maximum detail.
Solution Approach 2:
The patent introduces temporal dimension by using multiple frames and performing motion estimation across time. By analyzing sequences of frames rather than single images, the system can track object motion and improve accuracy for challenging cases where appearance varies significantly. This temporal dimension allows the DNN to learn motion patterns and compensate for appearance changes.
2Measurement precision
If classical methods are used for motion estimation, then accuracy is maintained for challenging cases, but computational efficiency deteriorates
Solution Approach 1:
The patent divides the image into multiple regions and applies different processing methods to different segments. This allows classical methods to be used only where necessary (challenging regions) rather than across the entire image, improving computational efficiency while maintaining accuracy where it matters most.
Solution Approach 2:
The patent applies the full computational power of classical methods only partially - specifically to regions identified as challenging through initial analysis or segmentation. For easier regions, simpler or DNN-based methods are sufficient, reducing overall computational load while maintaining accuracy for the difficult portions.
3Measurement precision
If object segmentation and matching are performed to improve motion estimation accuracy, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent uses a unified DNN architecture that performs multiple functions: segmentation, feature extraction, and motion estimation all within a single network. This multi-functionality reduces system complexity compared to having separate dedicated systems for each task, while still achieving high accuracy through the integrated approach.
Solution Approach 2:
The patent introduces feature vectors as an intermediary representation between raw image data and motion estimation results. These feature vectors capture essential object characteristics and facilitate matching between frames, improving accuracy while keeping the system manageable through this intermediate abstraction layer.
Data Source
AI summary
A method of generating a training dataset suitable for training machine learning algorithms to estimate the motion of objects, and for training a machine learning algorithm to perform motion estimation. A plurality of pairs of synthetic images are generated from obtained objects and backgrounds, each pair have a first frame and a second frame. The first frame includes a selection of objects in first positions and first orientations superimposed on a selected background, and the second frame includes the selection of objects in second positions and second orientations superimposed on the selected background. Also provided are processing systems configured to carry out these methods.