FBINeRF
Spanning Image

NeRF- A New Era for 3D reconstruction

  • FBINeRF focus on improving the current NeRF algorithms in terms of accuracy, speed, and also artifacts mitigating.
  • FBINeRF are supervised by SOTA depth estimation methods with both relative depth estimation and metric depth, and thus have a better generalization ability.
  • Flexible Bundle Adjusting can be used to establish more complicated camera poses model, including distortion, rotation, translation or even calibration. Compared to original NeRF, FBINeRF can deal with imperfect camera initialization and perform well in robustness.

Video Examples

BU Central Chapel
Garden
Pinecone
Stump

Research

This project is funded by VIP Lab at BU

Read More

Suggestion

Any suggestions please email:wuyifan1@bu.edu

Read More

Contact

Phone number: 617-870-9849

Read More

Point Cloud

Compared to traditional struction from motion method, FBINeRF spends quite less amounts of time for camera pose optimization. FBINeRF provides a continuous volumetric representation of a scene, allowing for high-resolution and accurate point cloud generation with fine details.NeRF captures both geometry and appearance information, resulting in more visually realistic point clouds that include color and shading details. FBINeRF can also model complex scenes, including scenes with non-Lambertian materials and varying lighting conditions, resulting in more versatile point cloud generation.

Read More

Flexible Bundle Adjusting

FBA can handle scenes with varying lighting conditions, partial occlusions, and challenging camera poses. It can refine camera poses and scene geometry more accurately even in cases where the initial estimates are not very accurate. It is also able to can refine intrinsic camera parameters (such as focal length, principal point, lens distortion) and extrinsic parameters (rotation and translation) simultaneously. This ensures the camera's calibration is accurate, which is crucial for achieving precise measurements and reconstructions.

Read More

FBINeRF-3D

Frequently Asked Questions

FBINeRF has a better generalization ability compared to other depth-supervised NeRF methods; while it can be used for further application with absolute value of depth estimation. Moreover, flexible bundle adjusting helps to enhence robustness of FBINeRF and allows to establish more complex models that close to situations in reality.

The increased complexity often leads to higher computational demands, making flexible bundle adjustment more resource-intensive and time-consuming. How to strike a balance and make most of flexible bundle adjustment stills needs to be discussed. Moreover, initializing non-rigid FBIormations and transformations accurately can be challenging, and incorrect initializations can lead to convergence issues or inaccurate results.

Our future work lies on adding anti-aliasing, simplifying algorithm complexity, and increasing training speed.