SIDAR

SIDAR是一个用于图像对齐和恢复的合成图像数据集，由柏林工业大学的计算机视觉与遥感实验室创建。该数据集通过3D渲染技术，将图像作为纹理添加到平面上，并引入多种光照条件、阴影和遮挡，以模拟真实世界场景。每个场景提供一系列扭曲图像及其对应的遮挡掩码、单应性和地面实况标签。SIDAR数据集适用于多种图像对齐和去除伪影的任务，如深度单应性估计、密集图像匹配、2D束调整、修复、阴影去除、降噪、内容检索和背景减法。此外，其渲染管道可定制，能作为现有数据集的数据增强，进一步提升特征学习能力。

SIDAR is a synthetic image dataset for image alignment and restoration, created by the Computer Vision and Remote Sensing Lab at the Technical University of Berlin. This dataset uses 3D rendering techniques to attach images as textures onto planar surfaces, and introduces diverse lighting conditions, shadows and occlusions to simulate real-world scenarios. Each scene provides a series of warped images along with their corresponding occlusion masks, homographies and ground-truth labels. The SIDAR dataset supports a wide range of image alignment and artifact removal tasks, including depth homography estimation, dense image matching, 2D bundle adjustment, inpainting, shadow removal, denoising, content retrieval and background subtraction. Furthermore, its customizable rendering pipeline can serve as a data augmentation tool for existing datasets to further enhance feature learning capabilities.

搜集汇总

背景与挑战

背景概述

SIDAR是一个由柏林工业大学创建的合成图像数据集，专为图像对齐和恢复任务设计。它通过3D渲染技术模拟真实场景，提供扭曲图像、遮挡掩码和地面实况标签，适用于深度单应性估计、修复和阴影去除等多种计算机视觉应用。此外，其可定制渲染管道支持数据增强，提升特征学习能力。