Related Experiment Video
Updated: Oct 19, 2025
06:54
Photorealistic Learned Landscapes for Augmented Reality
Published on: June 27, 2025
306
GigaMVS: A Benchmark for Ultra-Large-Scale Gigapixel-Level 3D Reconstruction
IEEE Transactions on Pattern Analysis and Machine Intelligence
|September 24, 2021
Summary
The GigaMVS dataset introduces gigapixel images for benchmarking large-scale 3D reconstruction algorithms. It reveals limitations in current multiview stereopsis (MVS) methods, highlighting needs for improved scalability and efficiency.
Area of Science:
- Computer Vision
- 3D Reconstruction
- Photogrammetry
Background:
- Multiview stereopsis (MVS) methods reconstruct 3D geometry and texture from images.
- Existing datasets lack comprehensive benchmarks for ultra-large-scale scenes with high-resolution details.
Purpose of the Study:
- Introduce GigaMVS, the first gigapixel-image-based benchmark for ultra-large-scale 3D reconstruction.
- Evaluate the scalability and efficiency of current MVS algorithms on challenging datasets.
Main Methods:
- Utilized gigapixel images offering wide field-of-view and high-resolution details.
- Collected ground-truth 3D geometry using laser scanning for scenes averaging 8667 m².
- Assessed state-of-the-art MVS methods using geometric and textural measurements.
Main Results:
- GigaMVS exposes scalability and efficiency issues in existing MVS algorithms.
- Analysis identified weaknesses in current methods for large-scale, complex scenes.
- Gigapixel images enable observation of both scene structure and fine local details.
Conclusions:
- GigaMVS provides a crucial benchmark for advancing 3D reconstruction.
- The dataset will drive the development of more robust, scalable, and accurate MVS algorithms.
- Future research should focus on improving algorithm performance for ultra-large-scale environments.
