COSD: A Challenging Object Segmentation Dataset


The COSD is a large-scale RGB-D dataset aiming to address the challenging object segmentation problem. This dataset contains 8,060 RGB-D images from 14 scenes with three semantic classes in terms of highly reflective objects, highly transparent objects and background. We use an iPad 2018 equipped with a Structure Sensor to capture these RGB-D images. For challenging objects, we consider common types of them (mainly including glass products, mirrors, plastic and metal) with different size, shape and color. For scenes, the COSD dataset contains 14 unique scenes (e.g. livingroom, bathroom, office, etc). Finally, we use Labelme( to annotate these images manually. Some demos in our COSD dataset are shown in Fig.1.
Fig.1. Demos of the COSD datset. The Depth image is colorized for visualization. The black, green and red in the semantic map represent the background, highly reflective and highly transparent objects respectively.


Download Link:

BaiDu Cloud:       Password: 6q88
Google Drive:
You can download the training set of the COSD via the proposed links. Afterdownloading the file, please send an email at to get the password of this file. The test set is not available now.
Attention please.
  • § The COSD dataset contains 8,060 RGB-D images (4,421 for training and 3,639 RGB-D images for test).
  • § Some depth values are missing due to the RGB-D sensor. The depth is saved as .png image in millimeter.
  • § We also provide the hha images caculated from the raw depth and camera parameters.
  • § In the semantic mask, 0, 1, 2 represent the background, highly transparent objects and highly reflective objects respectively.

Citing the dataset
Coming soon.



Guodong Zhang; Guijin Wang; Zhibin Xiao; Yang Li; Wenming Yang; Sifan Yang; Jing-hao Xue
Department of Electronic Engineering, Tsinghua University, Beijing 100084, China
For questions about the dataset, contact Guijin Wang:


This COSD dataset is made freely available to academic and non-academic entities for non-commercial purposes such as academic research, teaching, scientific publications, or personal experimentation.