Exploring Limits of Diffusion-Synthetic Training with Weakly Supervised Semantic Segmentation
This work addresses the problem of improving weakly supervised semantic segmentation for researchers and practitioners by offering incremental enhancements to existing diffusion-synthetic methods.
The paper tackled limitations in diffusion-synthetic training for semantic segmentation by introducing reliability-aware robust training, prompt augmentation, and LoRA-based adaptation, resulting in effectively closing the gap between real and synthetic training on datasets like PASCAL VOC, ImageNet-S, and Cityscapes.
The advance of generative models for images has inspired various training techniques for image recognition utilizing synthetic images. In semantic segmentation, one promising approach is extracting pseudo-masks from attention maps in text-to-image diffusion models, which enables real-image-and-annotation-free training. However, the pioneering training method using the diffusion-synthetic images and pseudo-masks, i.e., DiffuMask has limitations in terms of mask quality, scalability, and ranges of applicable domains. To overcome these limitations, this work introduces three techniques for diffusion-synthetic semantic segmentation training. First, reliability-aware robust training, originally used in weakly supervised learning, helps segmentation with insufficient synthetic mask quality. %Second, large-scale pretraining of whole segmentation models, not only backbones, on synthetic ImageNet-1k-class images with pixel-labels benefits downstream segmentation tasks. Second, we introduce prompt augmentation, data augmentation to the prompt text set to scale up and diversify training images with a limited text resources. Finally, LoRA-based adaptation of Stable Diffusion enables the transfer to a distant domain, e.g., auto-driving images. Experiments in PASCAL VOC, ImageNet-S, and Cityscapes show that our method effectively closes gap between real and synthetic training in semantic segmentation.