Adaptive Vision under Distribution Shift
Test-time adaptation, domain shift, calibration, robustness, and reliable model behavior at deployment.
Reliable Adaptive Vision Research
AI Researcher at Honda R&D Japan · PhD Student at The University of Tokyo
I work on reliable, adaptive, and deployment-aware computer vision systems for real-world autonomy.
I am an AI researcher at Honda R&D Japan and a PhD student at The University of Tokyo. My research focuses on adaptive and reliable vision systems, including test-time adaptation, continual and long-tailed learning, CLIP/VLM adaptation, and robust perception for AD/ADAS. I am especially interested in methods that remain reliable under distribution shift, class imbalance, and real-world deployment constraints.
Test-time adaptation, domain shift, calibration, robustness, and reliable model behavior at deployment.
Class-incremental learning, long-tailed continual learning, adaptive priors, and imbalance-aware recognition.
Training-free and lightweight adaptation of vision-language models, label propagation, graph-based adaptation, and robust open-vocabulary recognition.
Multi-camera perception, free-space and road segmentation, temporal consistency, long-tail driving conditions, lightweight deployment, and edge-aware perception systems.
A prior-correction framework for long-tailed continual learning that reduces imbalance-induced bias under incremental shifts.
View detailsAn empirical study of failure modes where spatial VLMs appear confident and consistent but ignore critical grounding evidence.
View detailsAdaptive prior correction for reliable long-tailed recognition in incremental scenarios.
Domain-resilient multi-camera segmentation pipelines designed for on-vehicle deployment constraints.
Occasional essays on research, discipline, travel, fitness, dharma, and life between India and Japan.
Visit Writing