Multi-Domain Object Detection with Teacher-Student Pseudo-Label Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional domain adaptation methods in object detection struggle to effectively learn the intrinsic features of objects across different domains, such as RGB, infrared, and thermal imaging, due to significant variations in object representation, leading to insufficient performance.
Innovation Solution
A multi-domain object detection method utilizing category-based domain learning, involving a teacher and student model, where the teacher model generates pseudo labels, and the student model learns through negative learning and unsupervised loss, with updates using an exponential moving average, to adapt to new domains.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional domain adaptation methods are used to train models on entire feature maps from source and target domains, then the model can process data from multiple domains, but the model fails to effectively learn intrinsic features of objects due to significant domain variations
Solution Approach 1:
The patent segments the feature map processing by introducing category-level discrimination. Instead of treating the entire feature map uniformly, the method divides the learning process into category-specific streams, where each category is processed separately through category-level discriminators. This segmentation allows the model to learn intrinsic features for each object category while being aware of domain differences, thereby improving detection accuracy across multiple domains.
Solution Approach 2:
The patent introduces category-level discriminators as intermediary components between the feature extraction layer and the detection head. These discriminators act as mediators that process category information and domain characteristics separately, enabling the model to distinguish between different categories while being invariant to domain variations. This intermediary mechanism resolves the contradiction by providing both multi-domain adaptability and reliable object detection through structured category-aware processing.
2Adaptability or versatility
If the model is trained using entire feature maps from both source and target domains, then domain adaptation can be performed, but the significant variations in object representation across domains (RGB, IR, thermal) prevent effective learning
Solution Approach 1:
The patent applies local quality by processing different categories of objects through separate streams with dedicated category-level discriminators. Each category stream learns the intrinsic features specific to that category while being adjusted for domain characteristics. This local processing approach ensures that each object category is represented accurately in its own feature space, maintaining measurement precision across different domains (RGB, IR, thermal) while preserving domain adaptation capability.
3Adaptability or versatility
If domain adaptation is performed using conventional methods on entire feature maps, then the model can handle multiple domains, but the model structure becomes overly complex and computationally intensive
Solution Approach 1:
The patent segments the complex domain adaptation problem into manageable category-specific sub-problems. By dividing the feature map processing into separate category streams, each handled by its own discriminator, the model achieves multi-domain processing ability without requiring a single monolithic complex structure. This segmentation reduces computational complexity by processing categories independently rather than simultaneously across all domains.
Solution Approach 2:
The patent applies partial action by processing only the necessary category information through category-level discriminators rather than processing every possible domain combination simultaneously. The model processes categories selectively and efficiently, achieving multi-domain adaptability through targeted category-aware processing rather than exhaustive domain-wide analysis, thereby reducing overall model complexity.
Data Source
AI summary
A multi-domain object detection method includes generating a teacher model and a student model from a pre-trained model, inputting an image with weak augmentation applied to a target image, for which an object is to be detected, to the teacher model, determining whether a pseudo label generated by the teacher model is below a preset threshold, performing negative learning for a class corresponding to the pseudo label when the pseudo label is determined to be below the threshold, inputting an image with strong augmentation applied to the target image to the student model, calculating an unsupervised loss by comparing a first prediction generated by the student model with the pseudo label, updating the teacher model using an exponential moving average (EMA) predetermined in the student model, and detecting an object in an image from another domain using the teacher model.