Trans10K

玻璃制成的窗户和瓶子等透明物体在现实世界中广泛存在。分割透明对象具有挑战性，因为这些对象具有从图像背景继承的不同外观，使得它们与周围环境具有相似的外观。除了这项任务的技术难度之外，只有少数以前的数据集是专门设计和收集来探索这项任务的，并且大多数现有数据集都存在重大缺陷。它们要么拥有有限的样本量，例如只有一千张没有手动注释的图像，要么使用计算机图形方法生成所有图像（即不是真实图像）。为了解决这个重要问题，这项工作提出了一个用于透明对象分割的大规模数据集，名为 Trans10K，由 10,428 张真实场景的图像组成，并带有仔细的手动注释，比现有数据集大 10 倍。 Trans10K 中的透明对象由于尺度、视点和遮挡的高度多样性而极具挑战性，如图 1 所示。为了评估 Trans10K 的有效性，我们提出了一种新的边界感知分割方法，称为 TransLab，它利用边界作为改善透明物体分割的线索。大量的实验和消融研究证明了 Trans10K 的有效性，并验证了在 TransLab 中学习对象边界的实用性。例如，TransLab 显着优于最近 20 种基于深度学习的对象分割方法，表明该任务在很大程度上尚未解决。我们相信 Trans10K 和 TransLab 对学术界和工业界都有重要的贡献，促进了未来的研究和应用。
Trans10K 数据集包含 10428 张图像，透明物体分为两类：（1）透明物体，如杯子、瓶子和玻璃，定位这些物体可以使机器人更容易抓取物体。 (2) 透明的东西？如窗户、玻璃墙和玻璃门。它可以让机器人学会避开障碍物，避免撞到这些东西吗？ 5000、1000 和 4428 幅图像分别用于训练、验证和测试。我们进一步根据难度将验证集和测试集分为简单和困难两部分。

Transparent objects such as glass windows and bottles are widespread in the real world. Segmenting transparent objects is challenging, as these objects inherit appearances from the image background, resulting in similar visual characteristics with their surrounding environment. Beyond the technical difficulties of this task, only a few prior datasets have been specifically designed and collected for this task, and most existing datasets suffer from significant drawbacks. They either have limited sample sizes—for example, only 1,000 unannotated images—or generate all images via computer graphics methods (i.e., not real-world images). To address this critical issue, this work proposes a large-scale dataset for transparent object segmentation named Trans10K, which consists of 10,428 real-world scene images with careful manual annotations, being 10 times larger than existing datasets. The transparent objects in Trans10K are highly challenging due to the high diversity of scales, viewpoints, and occlusions, as shown in Figure 1. To evaluate the effectiveness of Trans10K, we propose a novel boundary-aware segmentation method called TransLab, which leverages boundaries as a cue to improve transparent object segmentation. Extensive experiments and ablation studies demonstrate the effectiveness of Trans10K and validate the practicality of learning object boundaries in TransLab. For instance, TransLab significantly outperforms 20 recent deep learning-based object segmentation methods, indicating that this task remains largely unsolved. We believe that both Trans10K and TransLab make important contributions to both academia and industry, facilitating future research and applications.

The Trans10K dataset contains 10,428 images, and transparent objects are divided into two categories: (1) Common transparent objects such as cups, bottles, and glassware; localizing these objects can help robots grasp objects more easily. (2) Transparent obstacles such as windows, glass walls, and glass doors; this allows robots to learn to avoid obstacles and prevent collisions with these objects. 5,000, 1,000, and 4,428 images are used for training, validation, and testing, respectively. We further split the validation and test sets into simple and hard subsets based on difficulty levels.

搜集汇总

背景与挑战

背景概述

Trans10K是一个用于透明物体分割的大规模数据集，包含10,428张真实场景图像，并带有手动注释，比现有数据集大10倍，旨在解决透明物体因外观与背景相似而难以分割的挑战。数据集将透明物体分为两类（如杯子和窗户），支持机器人抓取和避障应用，并分为训练、验证和测试集，验证集和测试集进一步按难度划分，适用于计算机视觉语义分割研究。