Research
Usami Laboratory works on image processing, computer vision, and machine learning through four research themes.
Explore the research
Testing LLM Judge Reliability
We test three properties of LLM judges: output validity, non-tie preferences between identical answers (dark current), and selection of a specified target.
Anatomy-Preserving Augmentation via Generative AI
A workflow for augmenting medical training images with anatomical controls and generated-sample selection.
Separating Shape and Reflectance from a Single Image
A self-supervised method estimating scene properties like reflectance and shape from a single photo using relighting constraints.
Building IoT Prediction with Sparse Graphs and Physical Constraints
Spatiotemporal prediction that limits graph neighbors and incorporates bandwidth and connection constraints into training.
Mobile Extraction of Carotid Ultrasound IMT Regions
A conference abstract combining MobileNetV3 image classification with SegNet segmentation to examine efficient carotid ultrasound analysis.
Recommending Learning Actions from Activity Logs
A study of learning-action recommendations evaluated through changes in a grade predictor's outputs when activity logs are modified.
Image Translation for Polyp Detection Training
A study augmenting polyp-detector training with NBI-like images generated from white-light endoscopy.
Recovering Polyp Shape and Scale Using a Medical Suture
A study using a known-diameter suture and a lighting model to estimate approximate polyp shape and absolute scale from one endoscope image.
Polyp Shape Recovery with Two Lights and Normal Correction
A study combining a two-light photometric model with learned normal correction to estimate depth and shape from endoscope imagery.
Research Framework
We study methods for recovering shape, material properties, and motion from images and video, with applications in medicine, manufacturing, and infrastructure.
We combine optical, physical, and mathematical models with machine learning to examine the basis of an estimate and the conditions under which a method works.
Our generative AI research also examines image generation for training-data augmentation and the reliability of LLM-based evaluation.
Principles x Applications
The laboratory combines foundational modeling with applied research for real-world deployment.
Principle-oriented research
Image understanding grounded in optics, physics, geometry, and mathematics.
- Physics-based image analysis
- Spatiotemporal modeling
- Generative model evaluation
Application-oriented research
Technical development for concrete problems in medicine, industry, and society.
- Medical image diagnosis support
- Visual inspection
- Action recognition systems
Research Themes
Each theme introduces its research questions, approaches, and evaluation methods.
Medical Image Analysis, 3D Shape Reconstruction, and Diagnostic Support
Research on lesion detection, segmentation, 3D shape reconstruction, quantitative measurement, and diagnostic support using endoscopy, CT, cellular, and pathological images.
Physics-based Vision
Research on estimating shape, material, reflectance, and illumination from images by modeling reflection, transmission, scattering, and other optical phenomena.
Real-world Sensing and Spatiotemporal Modeling
Research on extracting temporal and spatial structure from videos and sensor data for action recognition, forecasting, and anomaly detection.
Generative AI, Learning Support, and Validation
Research on generative models, learning support, data augmentation, automatic question generation, and validation of generated outputs and LLM evaluators.
Evaluation Philosophy
We evaluate research outcomes from multiple viewpoints, not only by a single quantitative metric.
- Reproducibility: clear experimental conditions and traceable evaluation. - Generalization: robustness beyond a single dataset or imaging condition. - Interpretability: understanding why a method works and where it breaks. - Social impact: whether the result creates value in actual workflows.