HDR-DSP

训练数据集包含大小为128x128 (uint16) 的2660 SkySat L1A序列 (文件夹)。每个序列的长度从6到15帧不等。每个序列都是通过用整数翻译注册帧来生成的。这意味着帧故意不完全对齐，因为这将需要对数据进行重采样。
与每个序列相关联的有饱和掩模 () 和曝光比率 ()。帧的饱和掩码是布尔掩码，在样本有效的情况下等于1，否则为0 (由于饱和)。曝光率对应于Planet作为元数据给出的以毫秒为单位的每帧的曝光时间。通过将序列除以比率向量 (每帧一个浮点数)，所得序列的辐射测量应近似相等。但是，由于这些曝光时间的不精确性，这种均衡并不完美。

The training dataset consists of 2660 SkySat L1A sequences (folders) with a size of 128×128 (uint16). Each sequence has a variable length ranging from 6 to 15 frames. Each sequence is generated by performing integer translational registration on the frames. This means the frames are intentionally not perfectly aligned, as perfect alignment would require resampling the data.

Saturation masks () and exposure ratios () are associated with each sequence. The saturation mask for each frame is a boolean mask, which equals 1 when the sample is valid, and 0 otherwise (due to saturation). The exposure ratio corresponds to the exposure time per frame in milliseconds provided by Planet as metadata. By dividing the sequence by the ratio vector (one float per frame), the radiometric values of the resulting sequence should be approximately equal. However, this equalization is not perfect due to the imprecision of these exposure times.

搜集汇总

背景与挑战

背景概述

HDR-DSP是一个用于训练的数据集，包含2660个SkySat L1A序列，每个序列由6到15帧128x128大小的图像组成，通过整数平移注册生成，并附带饱和掩模和曝光比率以处理数据问题。该数据集由巴黎萨克雷大学和Kayrros于2022年发布，旨在支持多曝光推帧卫星的超分辨率研究。