T-ROBI: Transparent Reflective Objects In Bins Dataset

Jun Yang

Wenjie Xue

Sahar Ghavidel

Steven L. Waslander

,

This multi-view dataset targets transparent and reflective industrial objects.

Transparent Objects: Includes 2 industrial objects exhibiting varying levels of transparency and reflectivity.
Bin-Picking Scenes: Comprises 12 complex bin-picking scenes, where objects are randomly stacked.
Multi-View Acquisition: Captured from 55 distinct viewpoints using an EPSON C4L 6-axis robot.
Ground Truth Poses: Annotations of ground truth poses for each object in all scenes.
Ground Truth Depth: Offers high-fidelity depth maps captured with a high-end Ensenso camera, with the objects temporarily coated in anti-reflective scanning spray to ensure reliable depth measurements.
Synthetic Dataset: To facilitate network training, we created a large-scale synthetic dataset for T-ROBI objects.

EXPLORE MORE

If you're interested in reflective objects, please explore our sibling dataset, ROBI dataset, which specifically focuses on reflective object perception.

Anchor 1

REAL SCENES

SYNTHETIC SCENES

Anchor 2

LICENSE

The dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

PUBLICATIONS

Jun Yang, Wenjie Xue , Sahar Ghavidel, and Steven L. Waslander. Active 6D Pose Estimation for Textureless Objects using Multi-View RGB Frames. The International Journal of Robotics Research (Submitted), 2025.

Anchor 3

DOWNLOAD AND CODE

Synthetic Scenes

Python scripts to work on the T-ROBI dataset

Anchor 4

ACKNOWLEDGEMENTS

The creation of this dataset has been supported by Toronto Robotics and AI Laboratory and EPSON Canada Ltd.

Anchor 5

T-ROBI: Transparent Reflective Objects In Bins Dataset

Wenjie Xue

Sahar Ghavidel

,

,

,

This multi-view dataset targets transparent and reflective industrial objects.

Transparent Objects: Includes 2 industrial objects exhibiting varying levels of transparency and reflectivity.

Bin-Picking Scenes: Comprises 12 complex bin-picking scenes, where objects are randomly stacked.

Multi-View Acquisition: Captured from 55 distinct viewpoints using an EPSON C4L 6-axis robot.

Ground Truth Poses: Annotations of ground truth poses for each object in all scenes.

Ground Truth Depth: Offers high-fidelity depth maps captured with a high-end Ensenso camera, with the objects temporarily coated in anti-reflective scanning spray to ensure reliable depth measurements.

Synthetic Dataset: To facilitate network training, we created a large-scale synthetic dataset for T-ROBI objects.