Source-Free Detection With Strong Augmentation and Mean Teacher
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional unsupervised domain adaptive detection methods require labeled source data for adaptation, which is not feasible in real-world applications where data privacy or limited memory prevents data sharing, leading to performance degradation when detectors are tested on out-of-distribution data.
Innovation Solution
A method employing strong data augmentation in a pretraining phase to enhance a source-only model, followed by a self-trained mean teacher model in an adaptation phase using unlabeled target data, with confidence and consistency constraints to generate pseudo ground truth boxes for training a student model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional unsupervised domain adaptive detection methods are used, then detection performance on target domain is improved, but labeled source data must be accessible which prevents deployment in real-world applications with data privacy constraints
Solution Approach 1:
The invention extracts and removes the dependency on labeled source data from the adaptation process. The source data is completely excluded from the adaptation phase, with only the pretrained model weights being transferred. This allows the method to work solely with unlabeled target data while maintaining effective domain adaptation, resolving the contradiction between performance and data accessibility.
Solution Approach 2:
The invention introduces a mean teacher model as an intermediary that generates pseudo-labels for the target domain data. This intermediary mechanism enables the system to adapt to the target domain without direct access to labeled source data, by creating its own supervision signals through the teacher-student framework with confidence and consistency constraints.
2Adaptability or versatility
If source-free domain adaptive detection is implemented, then data privacy requirements are satisfied, but detection performance may degrade without access to labeled source data for adaptation
Solution Approach 1:
The invention enables the system to serve itself by generating its own training labels through the mean teacher model. The confidence constraint and consistency constraint allow the model to self-supervise on unlabeled target data, creating reliable pseudo-labels without external labeled source data. This self-service mechanism maintains high detection performance while satisfying data privacy requirements.
Solution Approach 2:
The mean teacher framework implements a feedback mechanism where the teacher model's predictions are used to train the student model, which in turn updates the teacher model iteratively. The confidence and consistency constraints provide feedback signals that guide this self-training process, ensuring that performance is maintained even without access to labeled source data during adaptation.
3Reliability
If strong data augmentation is applied to source images, then model robustness is improved, but training complexity increases
Solution Approach 1:
The invention applies strong data augmentation during the pretraining phase on source data before domain adaptation. This preliminary action builds robust feature representations that are domain-invariant, making the model more resilient to domain shifts. By establishing this robust foundation early, the subsequent adaptation phase becomes simpler and more effective.
Data Source
AI summary
A method for implementing source-free domain adaptive detection is presented. The method includes, in a pretraining phase, applying strong data augmentation to labeled source images to produce perturbed labeled source images and training an object detection model by using the perturbed labeled source images to generate a source-only model. The method further includes, in an adaptation phase, training a self-trained mean teacher model by generating a weakly augmented image and multiple strongly augmented images from unlabeled target images, generating a plurality of region proposals from the weakly augmented image, selecting a region proposal from the plurality of region proposals as a pseudo ground truth, detecting, by the self-trained mean teacher model, object boxes and selecting pseudo ground truth boxes by employing a confidence constraint and a consistency constraint, and training a student model by using one of the multiple strongly augmented images jointly with an object detection loss.