Semantic Segmentation Model Training Using Coarse Polygon Annotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semantic segmentation models require extensive fine pixel-by-pixel annotation, which is time-consuming and costly, and existing methods fail to effectively utilize coarse annotation data for training without introducing significant annotation noise.
Innovation Solution
A method and device for training a semantic segmentation model using coarse annotation data, employing unsupervised clustering and confidence-based reweighting to enhance the model's prediction accuracy by utilizing unannotated and incorrectly annotated pixels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pixel-by-pixel fine annotation is used, then prediction accuracy is improved, but annotation time and cost increase significantly
Solution Approach 1:
The patent segments the annotation process into two stages: coarse annotation at the polygon level and fine annotation at the pixel level. The coarse annotation provides initial segmentation of objects, while the fine annotation refines the boundaries. This segmentation allows the system to achieve high prediction accuracy without requiring entirely manual pixel-by-pixel annotation, thus reducing annotation time and cost.
Solution Approach 2:
The patent applies preliminary coarse annotation using polygon boundaries before performing fine pixel-level annotation. This preliminary action establishes the initial object regions, and subsequent fine annotation only needs to refine the boundaries within these pre-defined regions. This preliminary segmentation significantly reduces the time required for complete annotation while maintaining high accuracy.
2Loss of time
If coarse polygon annotation is used, then annotation time is reduced, but annotation noise and unannotated pixels increase
Solution Approach 1:
The patent implements a feedback mechanism where the coarse polygon annotation is used to generate initial predictions, which are then fed back into the system to guide the fine annotation process. The model learns from the coarse annotation and adjusts its predictions, creating a feedback loop that refines the annotation quality. This feedback mechanism allows the system to maintain high reliability while using time-efficient coarse annotation.
Solution Approach 2:
The patent changes the annotation parameter from pixel-level granularity to polygon-level granularity for coarse annotation, then dynamically adjusts the processing parameters during the fine annotation stage. By changing the parameter representation and processing methods based on the annotation level, the system achieves both time efficiency in coarse annotation and high reliability in the final results.
3Quantity of substance
If coarse annotation data is used alone, then annotation cost is reduced, but prediction accuracy decreases due to annotation noise
Solution Approach 1:
The patent merges coarse polygon annotation data with fine pixel-level annotation data in a unified training framework. The system combines the benefits of both annotation types by integrating them into a single model that processes both coarse and fine annotation information. This merging allows the system to achieve high prediction accuracy while maintaining the cost efficiency of using coarse annotation data, as both annotation types contribute to the training process.
Solution Approach 2:
The patent creates a composite annotation representation that combines coarse polygon boundaries with fine pixel-level details. This composite material approach allows the training model to process both the coarse structure and fine details together, achieving high prediction accuracy while using cost-effective coarse annotation data. The composite representation integrates the strengths of both annotation types without requiring entirely manual pixel-level annotation.
Data Source
AI summary
The present disclosure relates to a coarse annotation based semantic segmentation model training method and device operating by sending an original image into a semantic segmentation model for processing; acquiring a semantic feature map output by a specified convolution layer in the semantic segmentation model; sending the semantic feature map to a first training branch for training to obtain a first cross entropy loss value, wherein the first training branch is a fully supervised branch depending on a ground-truth map; sending the semantic feature map to a second training branch for training to obtain a second cross entropy loss value, wherein the second training branch is an unsupervised branch without a ground-truth map; and determining an overall loss function according to the first cross entropy loss value and the second cross entropy loss value.