Instance Segmentation Overlap Handling for Large Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional instance segmentation algorithms struggle with processing large images due to limited compute resources, and existing overlap handling techniques fail to accurately and reliably segment objects of arbitrary size across overlapping tiles.
Innovation Solution
A computer-implemented method that breaks down large images into tiles with overlapping regions, processes each tile using an instance segmentation algorithm, and selectively merges or discards instance boundaries based on an overlap degree metric to ensure accurate and reliable instance segmentation across the entire image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Volume of stationary object
If the image is broken down into tiles for processing, then the compute memory requirement is reduced and large images can be processed, but the segmentation accuracy in overlap regions deteriorates due to conflicting instance boundaries
Solution Approach 1:
The patent divides the large image into multiple overlapping tiles for parallel processing, reducing the memory burden on each processing unit. Each tile is processed independently by the instance segmentation algorithm, and the results are then merged by resolving conflicts in overlap regions through comparison of instance boundaries and confidence scores.
Solution Approach 2:
The patent implements a feedback mechanism where instance boundaries from overlapping tiles are compared and validated. Conflicting boundaries in overlap regions are resolved by evaluating confidence scores and geometric consistency, ensuring that the final merged segmentation maintains high accuracy despite the tiled processing approach.
2Productivity
If existing overlap handling techniques are used to merge tiles, then processing efficiency is maintained, but segmentation reliability deteriorates for objects of arbitrary size due to inaccurate merging strategies
Solution Approach 1:
The patent replaces simple geometric merging rules with a machine learning-based confidence score evaluation system. Instead of relying on fixed geometric criteria for merging instance boundaries, the system uses learned confidence scores to dynamically determine which boundaries to retain, improving reliability for objects of arbitrary sizes while maintaining processing efficiency.
3Measurement precision
If a single forward pass is used for small images, then segmentation accuracy is maintained, but processing capability deteriorates for large images due to limited compute memory
Solution Approach 1:
The patent segments large images into smaller overlapping tiles that can be processed in parallel within available compute memory constraints. Each tile undergoes the same instance segmentation algorithm as would be applied to a small image in a single forward pass, preserving local segmentation accuracy while enabling processing of arbitrarily large images through systematic tiling and merging.
Data Source
AI summary
An image is broken down into multiple tiles pairs of the tiles include an overlap region. The tiles can be processed by a instance segmentation algorithm. Techniques of overlap handling of multiple instance boundaries at least partly arranged in the overlap region are disclosed.