FruitBin: A tunable large-scale dataset for advancing 6D Pose estimation in fruit bin picking automation

Guillaume Duret, Mahmoud ALI, NICOLAS CAZIN, Alexandre Chapin, Florence Zara, Emmanuel Dellandrea, Jan Peters, Liming Chen

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Datasets and Benchmarks, 6D Pose estimation, Robotic, Bin Picking, Occlusion

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: This paper presents the largest 6D pose estimation dataset for fruit bin picking with benchmarking over scene generalization, camera generalization and occlusion robustness.

Abstract: Bin picking is a ubiquitous application spanning across diverse industries, demanding automated solutions facilitated by robots. These automation systems hinge upon intricate components, including object instance-level segmentation and 6D pose estimation, which are pivotal for predicting future grasping and manipulation success. Contemporary computer vision approaches predominantly rely on deep learning methodologies and necessitate access to extensive instance-level datasets. However, prevailing datasets and benchmarks tend to be confined to oversimplified scenarios, such as those with singular objects on tables or low levels of object clustering. In this research, we introduce FruitBin. It emerges as an unparalleled resource, boasting an extensive collection of over a million images and 40 million instance-level 6D poses. Additionally FruitBin differs with other datasets whith its inclusive representation of a wide spectrum of challenges, encompassing symmetric and asymmetric fruits, objects with and without discernible texture, and diverse lighting conditions, all enriched with extended annotations and metadata. Leveraging the inherent challenges and the sheer scale of FruitBin, we highlight its potential as a versatile benchmarking tool that can be customized to suit various evaluation scenarios. As a demonstration of this adaptability, we have created two distinct types of benchmarks: one centered on novel scene generalization and another focusing on novel camera viewpoint generalization. Both benchmark types offer four levels of occlusion to facilitate the study of occlusion robustness. Notably, our study showcases the difficulty of FruitBin dataset, with two baseline 6D pose estimation models, one utilizing RGB images and the other RGB-D data, across these eight distinct benchmarks. FruitBin emerges as a pioneering dataset distinguishing itself by seamlessly integrating with robotic software. That enable direct testing of trained models in dynamic grasping tasks for the purpose of robot learning. Samples of the dataset with its associated code are provided in the supplementary materials. FruitBin promises to be a catalyst for advancing the field of robotics and automation, providing researchers and practitioners with a comprehensive resource to push the boundaries of 6D pose estimation in the context of fruit bin picking and beyond.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9289