Trans2k

视觉对象跟踪主要集中在不透明对象上，而透明对象跟踪则很少受到关注。由于透明对象的独特性，它们的外观直接受到背景的影响，因此最近出现了第一个专用的评估数据集。我们通过提出第一个透明对象跟踪训练数据集Trans2k做出了贡献，该数据集由超过2k个序列组成，具有104,343个图像，并通过边界框和分割蒙版进行注释。注意到现代渲染器可以逼真地渲染透明对象，我们量化了特定于领域的属性，并渲染了包含现有对象训练数据集中未涵盖的视觉属性和跟踪情况的数据集。当使用Trans2k进行训练时，我们观察到在不同的现代跟踪架构集合中具有一致的性能提升 (高达16%)，并且由于缺乏适当的训练集而显示出以前无法实现的见解。数据集和渲染引擎将公开发布，以释放基于现代学习的跟踪器的功能，并在透明对象跟踪中培育新的设计。

Visual object tracking has primarily focused on opaque objects, while transparent object tracking has received far less attention. Owing to the unique characteristics of transparent objects, whose appearance is directly affected by the background, the first dedicated evaluation dataset for this task has recently been proposed. We contribute Trans2k, the first training dataset for transparent object tracking, which consists of over 2,000 sequences totaling 104,343 images, annotated with both bounding boxes and segmentation masks. Noting that modern renderers can photorealistically render transparent objects, we quantify domain-specific attributes and construct a dataset that encompasses visual attributes and tracking scenarios not covered by existing object tracking training datasets. When trained using Trans2k, we observe consistent performance improvements (up to 16%) across a diverse set of modern tracking architectures, and uncover previously unattainable insights that were previously limited by the lack of appropriate training data. The dataset and rendering engine will be publicly released to unlock the capabilities of modern learning-based trackers and foster novel designs in the field of transparent object tracking.

搜集汇总

背景与挑战

背景概述

Trans2k是首个透明物体跟踪训练数据集，包含2,000+序列和10万+图像，具有边界框和分割蒙版注释。该数据集通过现代渲染技术生成逼真透明物体图像，能显著提升跟踪算法性能(最高16%)，填补了该领域训练数据的空白。