Core Concepts
The author proposes a method called Information Transfer from the Built Geodesic Surface (ITBGS) to address limitations in few-shot generative model adaption. The approach involves creating a pseudo-source domain and utilizing interpolation and regularization to enhance image quality.
Abstract
The content discusses the challenges of generating images with limited data and introduces ITBGS, which consists of two modules: Feature Augmentation on Geodesic Surface (FAGS) and Interpolation and Regularization (I&R). The FAGS module creates a pseudo-source domain by projecting image features into the Pre-Shape Space, while the I&R module supervises interpolated images to improve quality. Experimental results demonstrate the effectiveness of ITBGS in achieving optimal results across diverse datasets in extremely few-shot scenarios.
Key points:
- Introduction of ITBGS for few-shot image generation.
- Description of FAGS and I&R modules within ITBGS.
- Demonstration of qualitative and quantitative experimental results.
- Comparison with other methods like StyleGAN2, FastGAN, and MixDL.
- Ablation studies on the proposed modules to evaluate their impact on image generation quality.
The proposed method shows promising results in balancing fidelity and diversity in generated images across various datasets.
Stats
Through qualitative and quantitative experiments, we demonstrate that the proposed method consistently achieves optimal or comparable results across a diverse range of semantically distinct datasets, even in extremely few-shot scenarios.
In recent years, there have also been some studies for image generation under few-shot setting.
Most of recent studies have explored model inversion to deduce the features of input real images.
Feature augmentation manipulates feature vectors, rather than augments only on the image level.
Some methods performed simple operations on features extracted by neural networks, such as adding noise and linear combination.
More complex transformations are also proposed for feature augmentation.
Instead of directly obtaining features, Mangla et al. leveraged self-supervision to obtain a suitable feature manifold before applying manifold mixup in their training procedure.
Quotes
"Finding the delicate balance between fidelity and diversity remains the top challenge in the field of extreme few-shot image generation."
"The proposed ITBGS produces commendable results across diverse 10-shot datasets."
"The trained generator can be used for further applications, such as few-shot image classification and instance segmentation."