Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

MGCFI-Net: Multi-scale globally aware feature learning with cross-view feature interaction for multi-view stereo.

Ming Han1, Hui Yin2, Aixin Chong3

  • 1State Key Laboratory of Advanced Rail Autonomous Operation, Beijing Jiaotong University, Beijing, 100044, China; Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, Beijing Jiaotong University, Beijing, 100044, China; School of Computer Science and Technology, Beijing Jiaotong University, Beijing, 100044, China.

Neural Networks : the Official Journal of the International Neural Network Society
|May 29, 2026
PubMed
Summary

Related Concept Videos

Multi-input and Multi-variable systems01:22

Multi-input and Multi-variable systems

Cruise control systems in cars are designed as multi-input systems to maintain a driver's desired speed while compensating for external disturbances such as changes in terrain. The block diagram for a cruise control system typically includes two main inputs: the desired speed set by the driver and any external disturbances, such as the incline of the road. By adjusting the engine throttle, the system maintains the vehicle's speed as close to the desired value as possible.
In the absence of...

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

[Efficiency of sperm donation: An analysis of 440 qualified sperm donors in Chongqing Human Sperm Bank].

Zhonghua nan ke xue = National journal of andrology·2020
Same author

The complete mitochondrial genome of the <i>Papilio paris</i> (Lepidoptera: Papilionidae).

Mitochondrial DNA. Part B, Resources·2020
Same author

Age and gender dependence of liver diffusion parameters and the possibility that intravoxel incoherent motion modeling of the perfusion component is constrained by the diffusion component.

NMR in biomedicine·2020
Same author

Acute invariant NKT cell activation triggers an immune response that drives prominent changes in iron homeostasis.

Scientific reports·2020
Same author

Development of a robust crystallization platform for immune receptor TREM2 using a crystallization chaperone strategy.

Protein expression and purification·2020
Same author

Diosmetin reduces bone loss and osteoclastogenesis by regulating the expression of TRPV1 in osteoporosis rats.

Annals of translational medicine·2020

MGCFI-Net improves 3D reconstruction by learning multi-scale features and interactions, enhancing accuracy in challenging, texture-lacking areas. This novel approach achieves state-of-the-art results on benchmark datasets.

Area of Science:

  • Computer Vision
  • 3D Reconstruction
  • Machine Learning

Background:

  • Learning-based multi-view stereo (MVS) methods show progress but struggle with accurate 3D reconstruction in weakly textured regions.
  • Existing methods often lack effective mechanisms for global contextual awareness and explicit cross-view feature interaction.

Purpose of the Study:

  • To introduce MGCFI-Net (Multi-Scale Globally Aware Cross-View Feature Interaction Network) for enhanced 3D reconstruction.
  • To address limitations in current MVS techniques, particularly in handling areas with limited texture.
  • To achieve state-of-the-art performance in 3D reconstruction quality and robustness.

Main Methods:

  • MGCFI-Net utilizes a Multi-Scale Hierarchical Perception (MSHP) module combining convolutions and Swin Transformer for discriminative feature extraction with global context.
Keywords:
Cross-view feature similarity supervisionFeature interactionMulti-scale hierarchical perceptionMulti-view stereo

Related Experiment Videos

  • Patch-expanding upsampling is employed for effective multi-scale feature fusion, preserving semantic information.
  • A Cross-View Feature Interaction (CVFI) module with epipolar attention mechanisms enhances feature matching between reference and source views.
  • Cross-View Feature Similarity Supervision (CFSS) enforces feature consistency via geometric warping for improved alignment and robustness.

    • Main Results:

    • MGCFI-Net demonstrates superior 3D reconstruction quality compared to existing methods.
    • The network achieves new state-of-the-art (SOTA) performance on widely recognized benchmarks like DTU, Tanks & Temples, and ETH3D.
    • Evaluations confirm the method's effectiveness and generalization capability, especially in challenging scenarios.

    Conclusions:

    • MGCFI-Net effectively addresses the challenge of 3D reconstruction in weakly textured regions.
    • The proposed network architecture and supervision strategy significantly advance the capabilities of multi-view stereo methods.
    • The results validate MGCFI-Net's potential for high-fidelity and robust 3D scene reconstruction.