Some links for published projects and talks.
FAST is an experimental annotation tool that focuses on mobile devices. We conducted user survey with annotation tasks.Learn more →
Toward Verifiable and Reproducible Human Evaluation for Text-to-Image Generation (CVPR2023, Co-authored)
Text-to-image generation is actually hard to evaluate. We showed that popular automatic metrics are often not correlated with human evaluation. We also proposed reliable manual evaluation protocol for text-to-image generation.Learn more →
Deploying and serving deep learning models focusing on performance is hard. Triton Inference Server is a great tool for deploying and serving deep learning models. I gave a talk about Triton Inference Server in GTC 2023 Japan AI Day.Learn more →