Booster Dataset

我们提出了一种新颖的高分辨率且具有挑战性的立体数据集，将室内场景与密集且准确的地面真相差异进行注释。我们的数据集特有的是几个镜面和透明表面的存在，即最先进的立体声网络故障的主要原因。我们的采集管道利用了一种新颖的深时空立体声框架，该框架可以轻松，准确地以亚像素精度进行标记。我们总共发布了在64个不同场景中收集的419个样本，并用密集的地面真相差异进行注释。每个样本包括高分辨率对 (12 mpx) 以及不平衡对 (左: 12 mpx，右: 1.1 Mpx)。此外，我们还提供手动注释的材料分割掩模和15k个未标记的样本。我们根据我们的数据集评估了最先进的深度网络，强调了它们在解决立体和未来研究的提示中的开放性挑战方面的局限性。

We propose a novel, high-resolution and challenging stereo dataset annotated with dense and accurate ground-truth disparity for indoor scenes. A distinctive feature of our dataset is the presence of several specular and transparent surfaces, which are the primary causes of failure for state-of-the-art stereo networks. Our acquisition pipeline leverages a novel deep spatio-temporal stereo framework, which enables easy and accurate annotation at sub-pixel precision. In total, we release 419 samples collected across 64 distinct indoor scenes, all annotated with dense ground-truth disparity. Each sample includes both high-resolution stereo pairs (12 megapixels) and unbalanced stereo pairs (left: 12 MPx, right: 1.1 MPx). Additionally, we provide manually annotated material segmentation masks and 15k unlabeled samples. We evaluate state-of-the-art deep networks on our dataset, highlighting their limitations in addressing the open challenges in stereo matching and providing insights for future research.

搜集汇总

背景与挑战

背景概述

Booster Dataset是一个高分辨率立体数据集，包含64个场景的419个样本，用于深度估计研究，特别关注镜面和透明表面的挑战。数据集由博洛尼亚大学2022年发布，提供密集的地面真相差异注释和材料分割掩模。