Raw2raw dataset

由于不同传感器品牌和型号的光谱灵敏度差异，相机传感器的原始 RGB 颜色会有所不同。本文重点研究不同传感器原始 RGB 颜色空间之间的映射任务。先前的工作使用成对校准来解决这个问题，以实现准确的颜色映射。虽然准确，但这种方法不太实用，因为它需要：（1）通过两个相机设备捕获一对图像，并在每个新场景中放置一个颜色校准对象； (2)对颜色校准对象进行准确的图像对齐或人工标注。本文旨在通过更实用的设置来解决原始空间中的颜色映射问题。具体来说，我们提出了一种半监督的 raw-to-raw 映射方法，该方法在一组成对的图像以及每个相机设备捕获的一组未成对的图像上进行训练。通过广泛的实验，我们表明，除了单校准解决方案之外，我们的方法与其他域适应替代方案相比取得了更好的结果。作为这项工作的一部分，我们从两个不同的智能手机摄像头生成了一个新的原始图像数据集。我们的数据集包括用于半监督训练和评估的未配对和配对集。
我们的数据集由两个不同的智能手机相机拍摄的未配对和配对图像组成：三星 Galaxy S9 和 iPhone X。未配对的图像集包括每个智能手机相机拍摄的 196 张图像（总共 392 张）。配对集包括 115 对用于测试的图像。除了这个配对集之外，我们还有另一组 22 个锚配对图像。

Due to differences in spectral sensitivity across sensor brands and models, the raw RGB colors captured by camera sensors vary. This paper focuses on the mapping task between raw RGB color spaces of different sensors. Previous works have addressed this problem using pairwise calibration to achieve accurate color mapping. Although accurate, this approach is impractical because it requires: (1) capturing a pair of images via two camera devices and placing a color calibration object in each new scene; (2) performing accurate image alignment or manual annotation on the color calibration object. This paper aims to solve the color mapping problem in raw spaces through a more practical setup. Specifically, we propose a semi-supervised raw-to-raw mapping method trained on a set of paired images and a set of unpaired images captured by each camera device. Through extensive experiments, we demonstrate that our method achieves better performance compared to other domain adaptation alternatives, in addition to single-calibration solutions. As part of this work, we generate a novel raw image dataset from two distinct smartphone cameras. Our dataset includes unpaired and paired sets for semi-supervised training and evaluation. Our dataset consists of unpaired and paired images captured by two different smartphone cameras: Samsung Galaxy S9 and iPhone X. The unpaired image set includes 196 images captured by each smartphone camera (totaling 392 images). The paired set includes 115 pairs of images for testing. In addition to this paired set, we have another set of 22 anchor paired images.

搜集汇总

背景与挑战

背景概述

该数据集旨在研究不同相机传感器原始RGB颜色空间之间的映射问题，提出了一种半监督的raw-to-raw映射方法。数据集包含由三星Galaxy S9和iPhone X拍摄的未配对图像392张和配对图像115对，用于半监督训练和评估。