The paper 'Real-Time Dental Panorama Generation from Handheld Intraoral Video', written by Kim Tae-gon (master's, first author), Jung Kang-hyun (master's), Eum Seung-ho (doctoral), Myung Jae-hong (master's), Lee Ji-hoon (master's) and Professor Park Un-sang (corresponding author) of the Computer Vision and Image Processing Laboratory (CVIP), has been accepted for publication at the outstanding international conference British Machine Vision Conference (BMVC) 2026 (acceptance rate: 406/1448 = 28.0%).

A handheld intraoral camera records tooth surfaces relatively cheaply and easily, which makes it well suited to clinical use. Its field of view is narrow, however, so the dentition is hard to inspect as a continuous structure and several local images must be examined separately. Technology that joins many frames into a single tooth-surface panorama is needed to make up for this, but generating a stable panorama is not easy when filming by hand.

Intraoral images in particular can destabilize ordinary feature-point image registration because of repetitive tooth shapes, weak enamel texture, specular reflection from saliva and lighting, and occlusion by the tongue, gums and cheeks. Existing SIFT/ORB-based methods and learning-based feature matching can produce incorrect correspondences, duplicated or missing teeth and ghosting in such conditions, and some learning-based methods require GPU computation and a separate model dependency.

With this in mind, the team proposes a real-time tooth-surface panorama generation framework that runs on the CPU without any separate training. Exploiting the temporal continuity of continuous video, it compares the current frame with subsequent candidate frames using ZNCC (zero-normalized cross-correlation), selects the most reliable frame, and builds left and right half-jaw panoramas sequentially through translation-based registration. This reduces reliance on unstable feature-point matching across repetitive tooth structures.

In the overlapping regions between frames, seam-aware blending finds a boundary where the brightness difference between the two images is small and blends smoothly around it. This is designed to reduce the seams, blurring and duplicated structures caused by registration error or differences in specular reflection, while preserving the local appearance of each tooth.

To join the two half-jaw panoramas filmed separately on the left and right, landmarks on the boundary teeth are extracted automatically. Tooth-centre landmarks are detected using the A-channel of the LAB colour space together with a distance transform, and the left and right panoramas are registered on that basis. This makes it possible to generate continuous tooth-surface panoramas for six view–jaw categories: the buccal, lingual and occlusal surfaces of the upper and lower jaws.

The proposed method was compared with a range of classical and learning-based stitching methods including ORB+RANSAC, SIFT+RANSAC and SuperPoint+LightGlue, and evaluated on visual quality, structural consistency, left–right panorama fusion performance and processing speed. It scored 4.45±0.71 on a five-point scale for overall visual quality and recorded the highest structural consistency at SSIM 0.8942±0.0283. It also achieved 48.12 FPS at 400×720, 34.33 FPS at 540×960 and 25.35 FPS at 600×1080, faster than the comparison methods at every evaluated resolution. Left–right panorama fusion also succeeded in most of the evaluated cases.

The research shows that the wide surface information of the dentition can be assembled into a continuous 2D panorama by exploiting the temporal continuity of handheld intraoral camera video, without a GPU-based learned matcher or any separate task-specific training. Online processing at around 30 FPS on mid-resolution input in particular points to the possibility of visualizing tooth surfaces close to real time on relatively light hardware.

Kim Tae-gon, the paper's first author and a master's student, said: "This research is significant in generating stable tooth-surface panoramas in the intraoral environment, where ordinary image stitching easily becomes unstable, by exploiting the temporal continuity of video without a separate learned model or a high-performance GPU. We aimed in particular to secure both visual continuity and processing speed, pointing to a direction usable in real handheld intraoral camera settings. Going forward we want to raise its applicability in real clinical settings through validation across a wider range of patients and filming conditions and through optimization for high-resolution processing. I am grateful to my co-researchers and to Professor Park Un-sang for his guidance."

The British Machine Vision Conference (BMVC), organized by the British Machine Vision Association (BMVA), is a major international conference in machine vision, image processing and pattern recognition. It is classified as an outstanding conference in AI and computer vision in the Korean Institute of Information Scientists and Engineers' list of outstanding conferences. The 37th BMVC 2026 will be held in Lancaster, United Kingdom, from 23 to 26 November this year.

References:

●           British Machine Vision Conference (BMVC 2026)

●           Website: https://bmvc2026.bmva.org/