real-colon-SfM
收藏资源简介:
REAL-Colon HF Triplets数据集是REAL-Colon数据集的一个重新打包版本,专为流式自监督AF-SfMLearner训练而设计。该数据集采用CC BY 4.0许可,面向深度估计任务,内容来源于结肠镜内窥镜视频,具有医学应用背景。数据以时间三元组的形式组织,每个数据行包含从原始视频帧中采样得到的前一帧(prev)、当前帧(curr)和下一帧(next)的RGB图像。此外,数据集还包括帧索引、序列标识、相机内参矩阵(K)、帧间实际步长(frame_stride)、源帧率(source_fps)和目标帧率(target_fps)等信息。由于原始REAL-Colon数据集不提供深度、遮挡或姿态的真值,因此深度图(depth)、遮挡图(occlusion)以及从前一帧/下一帧到当前帧的姿态变换矩阵(pose_curr2prev, pose_curr2next)字段均为空值,对应的布尔标志(has_depth, has_occlusion, has_pose)设置为false。数据集包含训练集(201,543个样本)、验证集(30,525个样本)和测试集(18,454个样本)三个标准分割,总数据量超过250GB。该数据集适用于自监督深度估计、光流、视觉里程计等计算机视觉任务在医学内窥镜场景下的研究与开发。
The REAL-Colon HF Triplets dataset is a repackaged version of the REAL-Colon dataset, specifically designed for streaming self-supervised AF-SfMLearner training. The original dataset is licensed under CC BY 4.0. It is oriented towards depth estimation tasks, with content derived from colonoscopy endoscopic videos, providing a medical application context. The data is organized in temporal triplets, where each row includes RGB images of the previous frame (prev), current frame (curr), and next frame (next) sampled from original video frames. Additionally, the dataset contains information such as frame indices, sequence identifiers, camera intrinsic matrix (K), inter-frame stride (frame_stride), source frame rate (source_fps), and target frame rate (target_fps). Notably, since the original REAL-Colon dataset does not provide ground truth for depth, occlusion, or pose, the fields for depth maps (depth), occlusion maps (occlusion), and pose transformation matrices from previous/next frames to the current frame (pose_curr2prev, pose_curr2next) are empty, with corresponding boolean flags (has_depth, has_occlusion, has_pose) set to false. The dataset includes three standard splits: training set (201,543 samples), validation set (30,525 samples), and test set (18,454 samples), with a total data volume exceeding 250GB. This dataset is suitable for research and development in self-supervised depth estimation, optical flow, visual odometry, and other computer vision tasks within the medical endoscopy scenario.




