PASCAL3D+
收藏资源简介:
3D 对象检测和姿态估计方法近年来变得流行,因为它们可以处理 2D 图像中的模糊性,并且与 2D 对象检测器相比,还可以为对象提供更丰富的描述。然而,大多数 3D 识别数据集仅限于每个类别的少量图像或在受控环境中捕获。在本文中,我们贡献了 PASCAL3D+ 数据集,这是一个用于 3D 对象检测和姿态估计的新颖且具有挑战性的数据集。 PASCAL3D+ 使用 3D 注释增强了 PASCAL VOC 2012 [1] 的 12 个刚性类别。此外,ImageNet [2] 为每个类别添加了更多图像。与现有的 3D 数据集相比,PASCAL3D+ 图像表现出更大的可变性,平均每个类别有 3,000 多个对象实例。我们相信该数据集将为研究 3D 检测和姿态估计提供丰富的测试平台,并将有助于显着推动该领域的研究。我们在我们的新数据集上提供了 DPM [3] 的变化结果,用于不同场景下的对象检测和视点估计,可用作社区的基线
Methods for 3D object detection and pose estimation have grown increasingly popular in recent years, as they can resolve ambiguities present in 2D images and offer richer object descriptions compared to 2D object detectors. However, most existing 3D recognition datasets are restricted to small quantities of images per category or are captured in tightly controlled environments. In this paper, we present PASCAL3D+, a novel and challenging dataset tailored for 3D object detection and pose estimation. PASCAL3D+ augments the 12 rigid categories from the PASCAL VOC 2012 benchmark [1] with 3D annotations. Additionally, we incorporate additional images per category sourced from ImageNet [2]. Compared to prior 3D datasets, PASCAL3D+ features greater variability in its image corpus, boasting over 3,000 object instances per category on average. We posit that this dataset will serve as a robust testbed for research on 3D detection and pose estimation, and will significantly advance the state of the art in this domain. We further provide results from a variant of DPM [3] for object detection and viewpoint estimation across diverse scenarios on our new dataset, which can function as a standard baseline for the research community.

- PASCAL3D+数据集首次发表,由Xiang等人在CVPR会议上提出,旨在结合PASCAL VOC和ImageNet数据集,提供3D对象的注释。
- PASCAL3D+数据集首次应用于计算机视觉研究,特别是在3D对象检测和姿态估计领域,推动了相关算法的发展。
- PASCAL3D+数据集被广泛用于深度学习模型的训练和评估,特别是在卷积神经网络(CNN)的背景下,提升了3D视觉任务的性能。
- PASCAL3D+数据集的扩展版本发布,增加了更多的3D注释和图像数据,进一步丰富了数据集的内容和多样性。
- PASCAL3D+数据集在自动驾驶和机器人视觉领域的应用显著增加,成为这些领域研究的重要基准数据集之一。
- 13D Object Representations for Fine-Grained CategorizationUniversity of California, San Diego · 2014年
- 23D ShapeNets: A Deep Representation for Volumetric ShapesPrinceton University · 2015年
- 3Multi-View Convolutional Neural Networks for 3D Shape RecognitionUniversity of Oxford · 2015年
- 4Learning to Segment Object CandidatesFacebook AI Research · 2015年
- 5Deep Learning for 3D Point Clouds: A SurveyUniversity of Science and Technology of China · 2020年



