Research

Research Areas

We connect human visual perception with computational models for images, videos, immersive media, and vision-language systems. Our work asks how visual content is experienced, measured, explored, and made reliable.

Reference turtle image without degradation

The same turtle image degraded by strong blur

Visual Quality Assessment

We develop perceptually aligned models that estimate the quality of images and videos as people experience them. This includes no-reference quality assessment, generative scanpath representations, and large multimodal models for image quality assessment.

Perceptual modeling · Image/video quality · LMM-based assessment

Human visual saliency map for the same panoramic scene

Visual Attention

We model not only where observers look, but also the sequence in which they explore a scene. These scanpath models help explain visual attention and support perceptual analysis for 360° images and videos.

Saliency · Scanpath prediction · Generative attention models

A vision-language model answering the same question for a clean image and a defocus-blurred image — **Clean versus defocus blur**The same visual question can produce an overconfident incorrect response after corruption.

Vision-Language Model Robustness

We examine how visual corruptions and perception bias affect vision-language models, then develop practical ways to diagnose failures and improve their use in perceptual tasks such as image quality assessment.

Corruption diagnosis · Perception bias · Reliable multimodal models