Keeping Segment Mask Quality with Self-generated Masks
Keywords: Formula-driven Supervised Learning, Generative Models, Semantic Segmentation
Abstract: Recent advances in generative models enable the creation of high-fidelity synthetic images from text prompts.
Using these images as training data for visual models holds promise for efficient learning in recognition tasks, such as segmentation, which require detailed annotation and are difficult to gather data for.
However, training on synthetic data might lead to model collapse in visual models. Effective methods are required to prevent this and achieve efficient learning with synthetic data.
This study focuses on semantic segmentation and investigates efficient learning techniques using synthetic data generated by generative models.
Specifically, we propose (i) a mask filtering method utilizing a segmentation model and (ii) a pre-training method named SemSegFDSL, which employs a dataset based on mathematical equations to construct visual models with a limited amount of synthetic data.
Experimental results on the Pascal VOC dataset demonstrate that our method improves performance by 7.8\% while using only half the synthetic data required by previous methods.
These findings suggest that the quality of teacher labels affects model collapse and that filtering based on teacher labels and pre-training methods with synthetic data can effectively prevent this issue.
Submission Number: 2
Loading