Continual Semantic Segmentation Model Training via Knowledge Distillation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional knowledge distillation-based continual semantic segmentation methods often cause confusion between background and novel classes, hindering the effective training of new models by failing to establish reliable class correspondence, which leads to information loss and poor generalization.
Innovation Solution
A method that classifies novel and background classes separately, predicts probabilities for each class, expands the class dimensions of old models to match current models, and calculates knowledge distillation loss to update the model without confusion, ensuring accurate knowledge transfer from existing to new models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional knowledge distillation methods are used for continual semantic segmentation, then the model can be updated with new data, but confusion between background and novel classes occurs due to unreliable class correspondence
Solution Approach 1:
The patent segments the class correspondence problem into two independent parts: novel class correspondence and background class correspondence. By separately handling these two types of correspondence, the method avoids the confusion that arises when treating them uniformly, thereby improving reliability while maintaining adaptability.
Solution Approach 2:
The patent applies different strategies for different parts of the problem: using knowledge distillation for novel class correspondence and using alternative methods for background class correspondence. This localized approach ensures that each part receives the appropriate treatment, improving overall reliability without sacrificing the model's ability to update with new data.
2Stability of the object's composition
If knowledge distillation is applied to maintain existing knowledge, then the new model output distribution is normalized to match the old model, but training of the novel class is hindered due to background-novel class confusion
Solution Approach 1:
The patent divides the knowledge distillation process into separate components: one for novel classes and one for background classes. This segmentation allows the output distribution to be normalized for stability while preventing background-novel class confusion from hindering novel class training productivity.
Solution Approach 2:
The patent applies knowledge distillation selectively - using it for novel class correspondence to maintain stability, while using different approaches for background correspondence to avoid interfering with novel class training, thus balancing stability and productivity.
3Ease of manufacture
If the class dimension of old models is not expanded to match current models, then computation is simplified, but accurate knowledge transfer cannot be achieved
Solution Approach 1:
The patent expands the class dimension of old model outputs to match the current model's class dimension. This dimensional expansion enables accurate knowledge transfer by ensuring that the old and new models operate in the same feature space, preventing information loss while maintaining computational feasibility through targeted expansion rather than complete retraining.
Data Source
AI summary
A method of training a knowledge distillation based continual semantic segmentation model may include training the continual semantic segmentation model based on training data, predicting a probability for each class of an output of a current continual semantic segmentation model that has been generated by being trained and a probability for each class of an output of an old continual semantic segmentation model, expanding a class related to the old continual semantic segmentation model so that a spatial dimension of a class related to the old continual semantic segmentation model is the same as a spatial dimension of a class related to the current continual semantic segmentation model, calculating knowledge distillation loss based on the predicted probability for each class, and updating the current continual semantic segmentation model based on the knowledge distillation loss.