Research Profile
Researcher and engineer at Sony Group Corporation, working on AI and 3D technology R&D. I collaborate with teams across Japan and abroad, with publications accepted at top-tier venues including CVPR and AAAI. Since 2025, I also serve as a board member of the Japan Committee for the International Olympiad in AI (JOAI).
Working on applied R&D at the intersection of 3D technology and video production. Collaborating with teams across multiple countries on shared research projects, with papers accepted at top venues including CVPR and AAAI.
Japan Committee for the International Olympiad in AI (JOAI)
Involved from the founding of the organization, working on domestic competition operations, sending Team Japan to the international event, and running AI outreach programs for middle and high school students. Also collaborating with MEXT and METI to help build a broader ecosystem supporting young AI learners in Japan.
B^3-Seg is a segmentation method for 3D Gaussian Splatting scene representations that balances boundary quality and training efficiency. By combining coarse region-level alignment with progressive boundary refinement, it produces stable, high-quality masks even in complex scenes while keeping practical compute costs. This work was accepted to CVPR 2026.
This method targets text-guided 3D-to-3D editing, transforming an input 3D scene into a new one according to natural-language instructions. It leverages a pretrained image-to-image diffusion model for per-view optimization while conditioning explicitly on the original 3D scene, improving 3D consistency and structural preservation. It also introduces dynamic scaling to control the strength of geometric deformation.
This work proposes a VAE built entirely with spiking neural networks (SNNs). To address the mismatch between conventional continuous latent sampling and spike-based computation, it introduces an autoregressive SNN design that models latent variables as Bernoulli processes, enabling variational learning on SNNs. Results on multiple datasets show generation quality comparable to or better than ANN-based baselines.