Hyeongju Lee 이형주
ITC-CSCC 2026 oral · BangkokITC-CSCC 2025 poster · SeoulCode public · MIT

Origin point optimization: sharper,
then faster volumetric reconstruction

Hyeongju Lee, Myungjin Cho, Min-Chul Lee (the 2025 paper also with Gilsu Yeo) · Kyushu Institute of Technology, Hankyong National University

9×9 elemental images81 → 1runsReconstructions needed to choose the origin
CPU · measured in the paper0.31sThe iterative search: an estimated 25.45 s (0.3142 s × 81)
GPU · CuPy, RTX 30900.0057sThe same calculation on a GPU

Fixed origin and optimized origin

A 1951 USAF resolution chart reconstructed from 9 × 9 elemental images, panels (a) and (d) of ITC-CSCC 2026 Fig. 6. Compare the edges of the bar groups.

USAF resolution chart reconstructed with a fixed origin USAF resolution chart reconstructed with the optimized origin Fixed originOptimized origin

Volumetric computational reconstruction (VCR) builds slices of a 3D scene by shifting elemental images, taken from many viewpoints, by an amount that depends on depth and overlapping them. The images can only move by whole pixels, which causes quantization error and blurs the result.

The 2025 paper reduced that error by optimizing the origin point that the overlap is measured from. But finding the best origin meant reconstructing many times, which was slow.

The 2026 paper computes this origin directly from the camera parameters alone, independent of the image. In the paper's experiments it gave the same origin and the same SSIM as the iterative search, and only the time dropped. This is not guaranteed in every case.

What VCR is

Diagram: a 3D object captured through a pinhole array as elemental images, back-projected through a virtual pinhole array to a reconstruction plane
Elemental images are back-projected through a virtual pinhole array to form a reconstruction plane at the chosen depth. ITC-CSCC 2026 Fig. 1

Method

The loop that reconstructed once per candidate origin became one calculation from the parameters.

Before: iterative search

  1. Acquire elemental images
  2. Pick a candidate origin
  3. Run VCR
  4. Score image quality and go back to step 2

Every candidate needs a full reconstruction, so the cost is high.

Proposed: direct calculation

  1. Acquire elemental images
  2. Enter the camera parameters and depth
  3. Compute the origin with the smallest quantization error from a formula
  4. Run VCR once

Only the parameters are used, not the image, so there is no loop. In the paper's experiments, image quality (SSIM) matched the iterative search.

Run time

9 × 9 elemental images. ITC-CSCC 2026 Tables III and IV.

Iterative searchCPU · estimated25.45 s
Direct calculation (proposed)CPU · measured0.31 s
Direct calculation (proposed)GPU · CuPy, RTX 30900.0057 s
81 reconstructions → 1The 25.45 s of the iterative search is an estimate: one reconstruction, 0.3142 s, times 81. The origin and SSIM were the same in the paper's experiments, which is not guaranteed in general.

Code

A Python release of both papers' methods: github.com/zoocasso/origin-point-optimization-vcr

  • Origin calculation (origin.py), VCR (vcr.py), quality metrics and a timing script
  • 21 tests, and 10,000 seeded cases checked against the re-implemented research formula
  • NumPy and OpenCV; CuPy is optional, for the GPU timing. MIT license

Cite

@inproceedings{lee2026accelerated,
  title     = {Accelerated Origin Point Optimization for
               Volumetric Computational Reconstruction},
  author    = {Lee, Hyeongju and Cho, Myungjin and Lee, Min-Chul},
  booktitle = {Proc. ITC-CSCC},
  year      = {2026}
}
@inproceedings{lee2025image,
  title     = {Image Enhancement of Volumetric Computational
               Reconstruction Using Origin Point Optimization},
  author    = {Lee, Hyeongju and Yeo, Gilsu and Cho, Myungjin and Lee, Min-Chul},
  booktitle = {Proc. ITC-CSCC},
  year      = {2025}
}
Contact

For roles and collaboration, please write by email.

I am looking for roles in data engineering, backend development and vision data pipelines.

Finishing my master's in March 2027