NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 26·5 min read
D3S2: Diffusion-Guided Dataset Distillation for Semantic Segmentation
TL;DR
Existing methods for dataset distillation primarily focus on image classification, neglecting dense tasks like semantic segmentation. D3S2, a Diffusion-guided Dataset Distillation framework, addresses challenges such as class imbalance and pixel alignment.
✦ Why It Matters
Engineers can leverage D3S2 to enhance dataset efficiency in semantic segmentation tasks, improving model performance with less data.
Key Takeaways
- Utilize D3S2 for efficient dataset distillation in segmentation tasks.
- Implement class-balanced mask selection to address class imbalance.
- Adopt diffusion-guided image synthesis for better spatial alignment.
- Incorporate segmentation-consistency and feature matching losses for improved training.
Full Summary
Dataset distillation (DD) compresses large datasets into smaller, synthetic sets while maintaining their training effectiveness, but it has been underutilized in semantic segmentation tasks. D3S2 introduces a two-stage framework to tackle key challenges: long-tailed class imbalance, strict pixel-wise alignment, and high computational costs.
The first stage, Class-Balanced Mask Selection, uses a greedy strategy to prioritize underrepresented classes, creating a representative mask set. The second stage, Diffusion-Guided Image Synthesis, employs a pretrained layout-to-image diffusion model to generate images based on these masks, ensuring spatial alignment.
To enhance the utility of the synthesized data, guided diffusion sampling incorporates a segmentation-consistency loss for pixel alignment and a class-wise feature matching loss for aligning feature statistics. Results indicate that at a compression rate of 1%, D3S2 achieves 24.99% and 35.49% mIoU on the ADE20K and COCO-Stuff datasets, respectively, outperforming random selection by 9.34% and 5.70%.
These findings suggest that D3S2 can significantly improve the efficiency of training models for semantic segmentation.