Deconvolutional Network Knowledge Transfer for Road Scene Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deconvolutional networks (DNs) for semantic segmentation in road scene analysis face inefficiencies in memory footprint and require large amounts of training data, leading to challenges in learning rare classes and generalizing to diverse road scenes, with existing datasets being limited and costly to annotate.
Innovation Solution
A source deconvolutional network (S-Net) is adaptively trained without memory constraints, and a target deconvolutional network (T-Net) with fewer parameters is developed to perform semantic segmentation, using a Multi-Domain Road Scene Semantic Segmentation (MDRS3) dataset that combines diverse training data with dense and sparse annotations to improve generalization and reduce memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deconvolutional networks are trained with large amounts of training data to improve segmentation accuracy, then manufacturing precision is improved, but loss of substance increases due to memory constraints and data scarcity
Solution Approach 1:
The patent applies preliminary action by pre-training a source deconvolutional network on large-scale datasets before constraining and adapting it for the target application. This allows the network to learn general road scene features beforehand, reducing the amount of additional training data needed for specific segmentation tasks while maintaining high accuracy.
Solution Approach 2:
The patent changes parameters by transitioning from an unconstrained source network to a constrained target network with reduced memory footprint. This parameter change allows deployment on embedded systems while maintaining segmentation accuracy through knowledge transfer from the pre-trained source network.
2Device complexity
If deconvolutional networks are constrained to reduce memory footprint for embedded deployment, then device complexity is reduced, but manufacturing precision deteriorates due to limited parameters
Solution Approach 1:
The source deconvolutional network is pre-trained on large-scale datasets before being constrained. This preliminary training embeds general road scene understanding into the network parameters, which are then transferred to the constrained target network, enabling accurate segmentation even with limited parameters suitable for embedded systems.
Solution Approach 2:
The patent creates a target deconvolutional network that copies the essential learned features and parameters from the source network. This copying process allows the constrained network to inherit the segmentation capabilities of the larger network while adapting to reduced memory constraints for embedded deployment.
3Adaptability or versatility
If diverse training data is used to improve generalization to diverse road scenes, then adaptability is improved, but loss of substance increases due to annotation costs
Solution Approach 1:
The source network is pre-trained on diverse large-scale road scene datasets beforehand, allowing it to learn generalizable features across different environments. This preliminary exposure to diversity reduces the need for extensive annotated data in the target application, improving generalization while reducing annotation costs.
Solution Approach 2:
The source deconvolutional network is designed to be universal by training on multiple domains and data types. This multi-functional training enables the network to handle diverse road scenes effectively, and this universality is transferred to the constrained target network, improving adaptability without requiring proportional increases in annotated training data.
Data Source
AI summary
A source deconvolutional network is adaptively trained to perform semantic segmentation. Image data is then input to the source deconvolutional network and outputs of the S-Net are measured. The same image data and the measured outputs of the source deconvolutional network are then used to train a target deconvolutional network. The target deconvolutional network is defined by a substantially fewer numerical parameters than the source deconvolutional network.