Advances in Camera Pose Estimation and 3D Scene Generation

The field of computer vision is witnessing significant advancements in camera pose estimation and 3D scene generation. Researchers are exploring novel approaches to estimate the relative position of cameras and objects, enabling applications such as autonomous navigation, robotics, and augmented reality. One notable direction is the development of methods that can handle sparse-view scenarios, where only a limited number of images are available. Additionally, there is a growing interest in using deep learning techniques to generate high-quality 3D models from 2D sketches or images. Noteworthy papers in this area include NeuroLoc, which proposes a neurobiological camera location method, and One2Any, which estimates the relative 6-degrees of freedom object pose using only a single reference-single query RGB-D image. Another significant contribution is the development of DiffusionSfM, a data-driven multi-view reasoning approach that directly infers 3D scene geometry and camera poses from multi-view images. These innovative approaches are pushing the boundaries of what is possible in computer vision and have the potential to enable a wide range of applications. Notable papers include: NeuroLoc, which enhances robustness in complex environments and improves pose regression. One2Any, which achieves state-of-the-art accuracy and robustness in 6D object pose estimation.

Sources

Relative position of a parabola or a hyperbola and an ellipse without computing intersection points

Are Minimal Radial Distortion Solvers Really Necessary for Relative Pose Estimation?

NeuroLoc: Encoding Navigation Cells for 6-DOF Camera Localization

T-Graph: Enhancing Sparse-view Camera Pose Estimation by Pairwise Translation Graph

PosePilot: Steering Camera Pose for Generative World Models with Self-supervised Depth

Learning Multi-frame and Monocular Prior for Estimating Geometry in Dynamic Scenes

A Birotation Solution for Relative Pose Problems

Corr2Distrib: Making Ambiguous Correspondences an Ally to Predict Reliable 6D Pose Distributions

seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models

EOPose : Exemplar-based object reposing using Generalized Pose Correspondences

One2Any: One-Reference 6D Pose Estimation for Any Object

S3D: Sketch-Driven 3D Model Generation

FastMap: Revisiting Dense and Scalable Structure from Motion

MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation

Convex Relaxation for Robust Vanishing Point Estimation in Manhattan World

DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion

3D Scene Generation: A Survey

Built with on top of