研究目的
To characterize human performance within the manifold of natural images using generative adversarial networks (GANs) to constrain parametric manipulations to remain within an approximation of the manifold of natural images.
研究成果
Human observers are sensitive to image manipulations that remain within an approximation of the manifold of natural images, as constrained by GANs. This sensitivity is related to changes in the configural structure of the images more than to changes in local image features. Observers are also sensitive to subtle aspects of paths along the model of the manifold, indicating a general tuning to natural images beyond detecting deviations from natural appearance.
研究不足
The images employed are relatively small (32x32 pixels), which may limit the generalizability of the findings to larger images. Additionally, the GAN's approximation to the manifold of natural images may not capture all aspects of natural images, particularly semantic content.
1：Experimental Design and Method Selection
The study used generative adversarial networks (GANs) to constrain parametric manipulations to remain within an approximation of the manifold of natural images. Two experiments were conducted: one to assess sensitivity to perturbations within the manifold and another to discriminate paths along the image manifold.
2：Sample Selection and Data Sources
Seven observers participated in the first experiment, and five in the second. Stimuli were samples from a GAN trained on the CIFAR10 dataset.
3：List of Experimental Equipment and Materials
Stimuli were presented on a Sony Triniton Multiscan G520 CRT monitor. A Minolta LS-100 photometer was used for monitor linearization.
4：Experimental Procedures and Operational Workflow
In the first experiment, observers performed a spatial two-alternative forced choice match-to-sample task. In the second experiment, observers discriminated between videos created by walking along paths in latent space.
5：Data Analysis Methods
Psychometric functions were estimated for each observer. Receiver operating curves (ROC) were calculated for predicting correct versus incorrect responses based on distance measures in latent or pixel space.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
