My name is Toma Okugawa. I’m a third-year student at the National Institute of Technology, Yuge College, a KOSEN on Yuge Island in the Seto Inland Sea. I grew up in the same town, Kamijima, so choosing the school was easy. Learning to code was not: the hook came in March of my first year, in 2025, while studying SQL for a certification exam. I started research that same year, and almost everything here has happened since.
Two threads
Two threads run through the work: making cameras see in the dark, and making phones understand human motion.
Most low-light enhancement today is deep learning running on a GPU, usually in the cloud. I wanted it to work in the opposite setting: a factory at night, a plant interior, a disaster site, where the hardware is a CPU and the video often cannot leave the site. So I went back to Retinex theory and built a series of training-free methods: a fast single-scale variant, then XCR, built during LEADING EDGE Shikoku for NOVA, a night-environment drone project, and then SD-Retinex, which fixed an instability in the math, not the code.
The second began with a teacher. In my second year, the teacher who kept pulling me into research was also the advisor of the kyudo club, so I built an app that runs pose estimation on a phone, detects kai (the full draw), and scores the archer’s form. When we presented it in 2025, the coaching text, generated by an external API, plainly lacked the knowledge to coach kyudo, so I started building my own models.
The constraint
Both threads share one rule: everything has to run on the device in front of you, with no cloud. That constraint turned out to be the business. I just found it the long way around.
In January 2026 I launched Second, an online programming school for elementary and junior-high students on remote islands like mine. I pivoted it in April and closed it in May. It did not fail on technology; it failed because I never knocked on a door, and it had nothing to do with my research. The two ventures that came after sit on top of research that already existed.
Beacon is what SD-Retinex was always for: a fully offline edge-AI system for inspecting dark environments such as factory floors. Kyudo AI is the kyudo method turned into a product: a smartphone-only coaching app that processes every frame on the phone.
What’s next
Two national finals in March 2027: Kigyoka Koshien with Beacon and the Lean Launchpad final with Kyudo AI. In October 2026 I’ll present different work at IEEE GCCE 2026, with Keito Inoshita of Kansai University, on re-defining Japan’s “vanishing municipalities” (my own town is one) with clustering and small language models.
When an interviewer asked what I would tell other KOSEN students, I said it is never too late to start. I started programming late, and I still got here. If you work on edge AI, low-light vision, or sports motion analysis, I’d like to hear from you on LinkedIn or GitHub.