Heng1987/armbench_defect_dataset

---
dataset_info:
features:
- name: rgb
dtype: image
- name: meta
struct:
- name: id
dtype: string
- name: label
dtype: string
- name: polygon
sequence:
sequence: int64
- name: sublabel
dtype: string
splits:
- name: train
num_bytes: 63572306115
num_examples: 141132
download_size: 70351144035
dataset_size: 63572306115
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: cc-by-4.0
---

This is data from the Amazon Armbench dataset (https://armbench.s3.amazonaws.com/index.html).

Each image is labeled with the failure mode {'nominal', 'package_defect', 'multi_pick'} in the 'label' field.
Failures are further specified in the 'sublabel' field {book_jacket', 'open_book_jacket', 'open_book', 'partial_box', 'empty_bag', 'torn_bag', 'open_box', 'crush_box'}.

Each image also contains a 'polygon' highlighting the area of interest.

To cite this dataset, please use

@article{mitash2023armbench,
title={ARMBench: An object-centric benchmark dataset for robotic manipulation},
author={Mitash, Chaitanya and Wang, Fan and Lu, Shiyang and Terhuja, Vikedo,
and Garaas, Tyler and Polido, Felipe and Nambi, Manikantan},
journal={arXiv preprint arXiv:2303.16382},
year={2023}
}