SlowTV

该数据集是从YouTube精心筛选出的大规模数据集，包含了来自徒步、驾车观光和潜水等多种环境下的长时事件视频中的170万张图片。这些视频以每秒10帧的速度提取，并且为了减少存储需求和避免自相似性，每250帧只保留了100帧连续画面。此外，该数据集还包含了通过COLMAP估计的相机内参信息。数据规模达到了170万张图片，其中110万张为自然风光，40万张为驾驶场景，18万张为水下画面。该数据集的任务是用于自监督的单目深度估计。

This large-scale dataset is carefully curated from YouTube, containing 1.7 million images extracted from long-duration event videos captured in diverse environments including hiking, scenic driving, and underwater diving. Frames are sampled at 10 frames per second, and to reduce storage requirements and avoid self-similarity, only 100 consecutive frames are retained per every 250 frames. Additionally, the dataset includes camera intrinsic parameters estimated via COLMAP. The total number of images in the dataset reaches 1.7 million, among which 1.1 million are natural landscape scenes, 400,000 are driving scenarios, and 180,000 are underwater footage. This dataset is intended for self-supervised monocular depth estimation tasks.