Proposal and Web Implementation of a Kyudo Form Evaluation Method Based on Real-Time Pose Estimation

In kyudo, Japanese archery, the decisive moment is kai: the full draw, held still for a few seconds before release. That is where form is judged, and it is also the one moment a camera can measure well, because the body is not moving. This paper proposed a way to evaluate kyudo form at that moment with real-time pose estimation running in a browser or on a smartphone, and described a web implementation of it.

I wrote it in my second year at Yuge College with two co-authors, Tomonari Masuzaki and Takahiro Makiyama, and presented it in September 2025 at the Shikoku-Section Joint Convention of the Institutes of Electrical and Related Engineers (SJCIEE 2025). The paper is in Japanese and appears in the refereed convention record. The convention is mostly attended by third- and fourth-year university students and KOSEN advanced-course students; I was the youngest presenter there, and the talk received the Outstanding Presentation Award from the convention’s executive committee.

Why I built it

The work started with a teacher rather than a research question. The teacher who kept pulling me into research was also the advisor of the kyudo club, so I built the first version partly to surprise him and partly so the club’s students could work on their form without waiting for a coach. A coach can watch each archer only briefly, and the part of the shot that matters lasts a few seconds.

Method

The system runs a pose estimator on live video, in the browser or on a phone. From the stream of joint positions it detects when the archer has settled into kai, then scores the form from the geometry of the joints at that instant. Written feedback for the archer was generated by sending the results to OpenAI’s API.

What it led to

The comment that mattered came after the talk: the coaching text came from an external API, and it plainly lacked the data and knowledge to coach kyudo. That is why I started building my own models instead of calling someone else’s.

The method itself became Kyudo AI, a smartphone-only app that watches you shoot, detects kai, scores your form and tells you what to fix, with every frame processed on the device so the video never leaves the phone. It was a finalist in the Youth Pitch Contest at the Startup World Cup 2026 Kyushu Regional in August 2026 and won the Grand Prize at Lean Launchpad Nagoya 2026 in September 2026.

It also tied my two research threads together. When I evaluated Fast Single-Scale Retinex, my first low-light enhancement method, the practical test was whether a pose estimator could still find fingertips and joints in dark footage after enhancement. Both lines of work share one constraint: everything has to run on the device in front of you, with no cloud.