"OpenPack: A Large-scale Dataset for Recognizing Packaging Works in IoT-enabled Logistic Environments" (PerCom2024 - Presentation)

This is a slide for the presentation at PerCom 2024 (March 13).

== Citation ==
Naoya Yoshimura, Jaime Morales, Takuya Maekawa, Takahiro Hara, “OpenPack: A Large-scale Dataset for Recognizing Packaging Works in IoT-enabled Logistic Environments”. Proceedings of IEEE International Conference on Pervasive Computing and Communications (2024).

== Links ==
PerCom2024: https://www.percom.org/full-program/
OpenPack HP: https://open-pack.github.io/

Transcript

1. N.Yoshimura, J. Morales, T. Maekawa, T.Hara (Osaka University, Japan) [PerCom2024, March 13] OpenPack: A Large-scale Dataset for Recognizing Packaging Works in IoT-enabled Logistic Environments
2. OpenPack Dataset Background CPS / Digital Twin Manager Factory ML ML’ Worker’s Activity Data IoT Device Sensor Readings In Smart Factories • Many sensors and IoT devices are installed. • Some works mainly depend on human workers. Supportive techniques to track worker activities are required. ⇒ Work Activity Recognition 2
3. OpenPack Dataset Challenges of Work Activity Recognition Dataset • Most of the public dataset focuses on the ADL (activity of daily living). • Public datasets in industrial domains containing complex activities is limited. Lack of datasets for industrial domains
 • Datasets for manual task provide only vision-related modalities. Limited Modality
 Lack of Io Data & Metadata
 • A lot of information related to the work activity is available in the system in the factory. ◦ e.g.) Order Management System → Items to be packed, • Existing activity recognition datasets do not provide these data. 3
4. OpenPack Dataset OpenPack Dataset • New large-scale multimodal activity recognition dataset in industrial domain. • Designed for developing activity recognition models with IoT-enabled devices. 0100
 Picking
 0200
 elocate Item Label 
 0300
 Assemble Box
 0400
 Insert Items
 0500
 Close Box
 0600
 Attach Box Label
 0700
 can Label
 0800
 Attach hipping Label 
 0900
 Put on Back able
 1000
 Fill out Order
 Work Operation & Action
 Period
 201
 emove Item Label
 204
 Write Check mark
 4 https://youtu.be/RiZ7kVpIHwU?si=DipjAt2s0Kp8KN6q
5. OpenPack Dataset OpenPack Dataset Recordings 53.8H Subjects 16 Annotations Work Operation Action 20K 53K 5 Modality 9+IoT Metadata
6. OpenPack Dataset Public Datasets for Manual Works
 Subjects 16* 14 Work Periods 2048 (~100 /person) ~30 trial/person Recording Length 53.5h 12.5h Activity Class 10 / 32* 8* / 19 Modality Depth + Keypoints +Acc +Gyro +Ori +LiDAR +EDA + BVP +Tmp Keypoints +Acc +Gyro Metadata / Subject Available Available ー OpenPack LARa [2020] 12 5 times / person 25h 4 / 17 Acc+Gyro+Mag +Mic+Loc +Object+Ambient - Opportunity [2010] ** work operation / action 9 (No Repetition) 10h 18 Acc+Gyro +Mag+Ori +HR+Temp Available PAMAP [2012] Metadata / System Available No No No IoT Data Available No (Object Sensor) No 6 Annotation Labels 20,161 / 53,286** N/A 2640 / 7153 * N/A *Not work related classes. * 12 subjects have work experience. *Locomotion/Hand interaction occurrence
7. OpenPack Dataset Data Collection Scenarios • 5 sessions/subjects, 1 session = 20 periods • The difficulties in packaging work recognition depend on the various factors. ⇒ Four scenarios are prepared to incrementally solve challenges. # of Items 54 Work Procedure Follow the work instruction as much as possible Scenario 1 Irregulars No Alarm Sound No 75 Able to alter the work procedure at subjects decision. Scenario 2 No No 75 Able to alter the work procedure at subjects decision. Scenario 3 Yes No 75 Able to alter the work procedure at subjects decision. Scenario 4 Yes Yes Ideal Wild 7
8. OpenPack Dataset Environment & Sensors Front-view • Kinect (Depth + Keypoints) • LiDAR Top-view • RealSense (Depth) Vision Modality Wearable ensors • IMU x 4 (Acc + Gyro + Quaternion) • Empatica E4 x 2 (Acc + BVP + EDA + Temp) Io Devices • Handheld Scanner • Label Printer 1 1 2 3 2 3 8
9. OpenPack Dataset Metadata Two types of metadata are available. ubject-related Metadata
 Order Sheet Order-related Metadata
 1 2 • Work Experience • Dominant Hand • Gender, Age Online order management system can provide information of a set of items to pack in the order. Item List Subject Data 9 Metadata can be used for recognition. ⇒ Useful for estimating the compositions of the work. (e.g. # of items = repetition of “remove labels” action.)
10. OpenPack Dataset IoT Data Advantage • Strong connection between devices in use and worker’s activity. ⇒ High-confidence source IoT Data can
 Item Label
 can
 Printer
 Acc can Label Operation Disadvantage • Data is generated only when a worker operates the devices. ⇒ Sparse data Existing sensor fusion techniques assume normal sensor data such as acceleration. ⇒ Method to make the best use of this high-confidence but sparse data source can enhance work activity recognition performance. 10
11. OpenPack Dataset Analysis: Factors Impacting Processing Time Length of Work Operation • The # actions performed in one period differs for each order. ◦ “Relocate item label” ⇒ 1+ action/item ◦ “Attach Box label” ⇒ 1 action/period # of Item, Box Size • The more items, the longer it takes. • The bigger the box, the longer it takes. OpenPack contains a huge variation of data! 11
12. OpenPack Dataset Benchmark • Task: 10 work operation recognition at 1 Hz • Input: Acceleration data from left wrist (window size = 60s) • Metrics: F1-measure (macro average) (Note: Protocols differs for each benchmark scenario.) Evaluation Protocol
 Baseline Models
 • CNN [F. Ordonez 2016] • U-Net [Y. Zhang 2019] … CNN-based segmentation model • DeepConvLSTM (DCL) [F. Ordonez 2016] … CNN + LSTM • DCL + Self-attention [S.P. Singh 2021] … Self-attention • ConformerHAR [ Y.-W. Kim 2022] … Transformer-based model • LOS-Net(-R) [N. Yoshimura 2022] … Designed for manual work recognition model 12
13. OpenPack Dataset Benchmark 1: Data-rich Setting Activity recognition with the enough amount of training data. • Objective: Confirm the upper bounds of recognition performance. • Protocol: Leave-one-subject-out CV Results • LOS-Net(-R) achieved 0.83 in Scenario 1. • Scores for Scenario 4 (Rushed) was lower than others. ⇒ Models are not speed invariant. • Contents of order have impact on performance. ⇒ Models are not robust to changing orders. Development of model that is robust to the work speed and orders are necessary. 13
14. OpenPack Dataset Benchmark 2: Data-scarce Setting Activity recognition with the limited amount of training data • Objective: More realistic than Scenario.1 • Training: Data from the 3rd session only (= 20 periods ~= 5h annotation.) • Test: Data from the remaining session. Results • Benchmark 1 ⇒ Benchmark 2 @ Scenario 1 ◦ LOS-Net(-R): 0.83 ⇒ 0.67 (- 0.16pt↓) ◦ Conformer : 0.78 ⇒ 0.53 (- 0.24pt↓) • CNN outperformed U-Net. • DCL outperformed DCL with Self-attention. 14 Technique to train the SOTA models with the limited training data is necessary.
15. OpenPack Dataset Research Directions with OpenPack Dataset • Metadata-aided activity recognition • Fusion with high-confidence data source. • Speed-invariant activity recognition Others Topics • Transfer learning across subjects, across modalities. • Skill assessment using sensor data and metadata • Counting the number of necessary actions / packed items. • Estimating worker’s level of fatigue using sensor and physiological data. • Detecting mistakes and accidents in the work process. Metadata IoT 15 Metadata IoT 4 Scenarios 16 Subjects 9 Moldaity 16 Subjects Metadata Action Label EDA+BVP Action Label+ Related Freatures Related Freatures
16. OpenPack Dataset Summary • The largest multimodal work activity dataset of packaging work. ◦ 53H+ Recording, 16 Subjects, IoT + Metadata ◦ 20K Work Operation Labels, 53K Action Labels • For more information ⇒ Visit our Website ! • Check sample data ⇒ GitHub • Try it out ⇒ Preprocessed Data at Zenodo. OpenPack Dataset
 Dataset is Available Now! 
 16 Preprocessed Data Label (Work Operation) IMU (Acc) Website