Object Detection Network Domain Adaptation With Pseudo-Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using CNNs face challenges in adapting to new domains with limited or no labeled data, particularly in one-stage type detectors, and there is a need to improve performance by addressing domain gaps and pseudo-label generation.
Innovation Solution
A learning apparatus and method that employs domain adaptation through supervised learning, adversarial learning, and self-learning to generate pseudo-labels, using integrated and manifold synthesis pseudo-labels, and computes total losses to update detection networks, incorporating techniques like gradient reversal layers and focal loss to align feature distributions across domains.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If domain adaptation is performed using weak supervised learning without teaching labels, then the applicability to new domains is improved, but the measurement precision of object detection positions deteriorates
Solution Approach 1:
The patent introduces pseudo-labels as an intermediary element that bridges the gap between unlabeled target domain data and the detection network. These pseudo-labels, generated through self-learning mechanisms, serve as intermediate teaching signals that enable the network to adapt to new domains without requiring actual annotated data, thus resolving the contradiction between adaptability and precision
Solution Approach 2:
The patent implements a feedback mechanism where the detection network generates pseudo-labels from its own predictions, which are then used to retrain and improve the network. This self-feedback loop allows the system to continuously refine its performance in new domains, maintaining measurement precision while achieving broad adaptability through weak supervised learning
2Measurement precision
If multiple loss functions are computed and combined, then the learning accuracy is improved, but the device complexity increases
Solution Approach 1:
The patent merges multiple loss functions (supervised loss, adversarial loss, and self-learning loss) into a unified total loss computation. By combining these diverse optimization objectives into a single comprehensive loss function, the system achieves improved learning accuracy through multi-objective optimization while managing computational complexity through integrated rather than separate processing
Solution Approach 2:
The total loss function serves multiple functions simultaneously: it supervises detection accuracy, aligns domain distributions through adversarial learning, and refines pseudo-labels through self-learning. This multi-functional loss computation improves learning accuracy across multiple dimensions without requiring separate complex processing systems for each objective
Data Source
AI summary
According to one embodiment, a learning apparatus includes processing circuitry that generates a detection network for detecting an object from an image belonging to a target domain. The processing circuitry computes a detection task loss by supervised learning of the detection network. The processing circuitry computes a domain identification loss by adversarial learning of the detection network. The processing circuitry computes a self-learning loss by self-learning of the detection network. The processing circuitry computes a total loss based on weighted addition of the domain identification loss, the detection task loss and/or the self-learning loss. The processing circuitry updates learning parameters of the detection network by minimizing the total loss.