---

Dataset Generation

The dataset was generated using the Blender 3D software with an emphasis on generality and diversity. Over 500 high-definition backgrounds (HDRI) were used, providing a wide variety of natural illumination and environments. In addition, over 13,000 random objects were used for both the background and the vessel content. Finally, over 1400 material textures (PBRs) were used for the ground plane. The vessels were procedurally generated with an unlimited number of different curves, shapes, and materials. Two types of content were generated inside the container. The first type is random objects taken from the ShapeNet dataset and put inside the vessel. The second content type was liquids with various properties simulated using the Blender MantaFlow tool with effects such as splashing, foam, and bubbles. Container shapes and materials for both vessels and content were procedurally generated. Altogether, this makes the TransProteus dataset one of the most diverse synthetic datasets in terms of environment, illumination, materials, objects, and setting. In addition, we created a small dataset containing 104 real-world photos with depth maps of both the vessel and its content; this set was created using the RealSense depth sensor.

Contact

The LabPics Dataset was created by the Matter Lab, Alan Aspuru group, feel free to contact us using this email.

Mirror Dataset Download

Full Dataset Mirror 1 Full Dataset Mirror 2 Sample of Dataset 1 Sample of Dataset 2