I-BADAS Dataset: Intralogistics Bin Anomaly Detection And Segmentation

Authors

Jose Moises Araya-Martinez¹², Sarvenaz Sardari², Katharina Engel², Pablo Rey Valiente², Madhan Raj Gopi Akila², and Jens Lambrecht³

¹ Industrial Automation Technology, Technical University Berlin, 10587 Berlin, Germany
² Future Manufacturing Technologies, Mercedes-Benz AG, 71063 Sindelfingen, Germany
³ Institute for Cognitive Robotics, Technical University Braunschweig, 38106 Braunschweig, Germany

Overview

The dataset accompanies our paper: I-BADAS: A Multi-Modal Dataset for Anomaly Detection with Co-Occurring Anomaly Types in Variable Intralogistics Settings.

The I-BADAS Dataset (Intralogistics Bin Anomaly Detection And Segmentation) is a comprehensive multi-modal benchmark designed for anomaly detection in intralogistics settings. This repository provides:

1. The complete I-BADAS dataset: 3,156 RGB-D images of industrial containers captured under variable industrial conditions such as lighting, camera to object position, background and distractors.
2. Multi-modal sensor data: Synchronized RGB and depth data acquired using Zivid 2+ M60 as industrial-grade with 0.24 mm spatial resolution and greater than 99.8% dimensional trueness and Realsense-D435 as consumer-grade RGB-D cameras having depth accuracy less than 1% of the measured distance (approximately 2.5–5 mm error at 1 m) with corresponding camera settings and information of the scene.
3. CAD models and digital twin assets: Accurate CAD representations of all container types for sim-to-real transfer, plus high quality scan of boxes using Faro Quantum X.S scanner.
4. Annotations and ground truth: Segmentation masks of boxes, and anomalies, annotations in Coco format.

Citation

If you use this dataset, please cite:

@article{arayamartinez_ibadas,
  title   = {I-BADAS: A Multimodal Dataset for Anomaly Detection with Co-Occurring Anomaly Types in Variable Intralogistics Settings},
  author  = {Araya-Martinez, Jose Moises and Sardari, Sarvenaz and Engel, Katharina and Rey Valiente, Pablo and Gopi Akila, Madhan Raj and Lambrecht, Jens},
  note    = {Submitted for Review},
  year    = {2026}
}

Dataset Statistics

Dataset Summary

Property	Value
Total Images	3,156
Training Images Task 1	1,003
Training Images Task 2	714
Test Images	1,439
Number of Container Types	3
Modalities	RGB + Depth (RGB-D)
Cameras	Zivid 2+ M60, Intel RealSense D435
Annotation Types	Segmentation Masks, 6D Pose, Image-Level Labels
Tasks	Emptiness Detection, Cleanliness Detection
CAD Models Included	Yes
Sim-to-Real Support	Yes

Image-Level Label Definition

Label	Description
empty_clean	Empty container with no contamination or residual content
empty_dirty	Empty container containing one or more nuisance contaminants
non_empty	Container containing residual objects, regardless of additional contaminants

Binary Anomaly Classes & Distribution of Test Set

Image Category	Image Count	Binary Class	Class	No. Instances
empty_clean	524	Anomaly-free	box	524
			permanent_mortise	1,178
non_empty	319	Anomalous	residual_content	1,688
empty_dirty	596	Anomalous	nuisance_sticker_inside	2,480
			nuisance_oil	1,190
			nuisance_gunk	644
			nuisance_color	799
			nuisance_paper	262
			nuisance_sticker_outside	127
			nuisance_plastic	106
			nuisance_dent	93
			nuisance_dust	84
			nuisance_foil	64
Total Images	1,439	—	—	—

Figure 1: Segmentation mask distribution per anomaly detection class.

Figure 2: Image distribution per scene and anomaly detection task.

Figure 3: Binary Anomaly Classes & Distribution.

Folder Structure

i-badas
├── models
│   ├── 3D_Scans
│   │   ├── KLT_4314.glb
│   │   ├── KLT_4315.glb
│   │   ├── partial_scans
│   │   │   ├── KLT_4314
│   │   │   ├── KLT_4315
│   │   │   └── Set_Box
│   │   └── Set_Box.glb
│   ├── KLT_4314.glb
│   ├── KLT_4314.ply
│   ├── KLT_4314.stl
│   ├── KLT_4315.glb
│   ├── KLT_4315.ply
│   ├── KLT_4315.stl
│   ├── models_info.json
│   ├── Set_Box.glb
│   ├── Set_Box.ply
│   └── Set_Box.stl
├── README.md
├── test
│   ├── 000001
│   │   ├── annotations.coco.json
│   │   ├── anomaly
│   │   │   ├── depth
│   │   │   ├── ground_truths
│   │   │   ├── mask_all
│   │   │   ├── mask_nuisance
│   │   │   ├── mask_residual
│   │   │   ├── mask_visib
│   │   │   ├── point_clouds
│   │   │   ├── rgb
│   │   │   ├── scene_camera.json
│   │   │   └── scene_info.json
│   │   ├── good
│   │   │   ├── depth
│   │   │   ├── ground_truths
│   │   │   ├── mask_visib
│   │   │   ├── point_clouds
│   │   │   ├── rgb
│   │   │   ├── scene_camera.json
│   │   │   └── scene_info.json
│   │   └── poses.json
│   ├── 000002
│   ├── 000003
│   ├── 000004
│   ├── 000005
│   └── 000006
└── train
    └── real
        ├── task1
        │   ├── 000001
        │   │   ├── rgb
        │   │   ├── point_clouds
        │   │   ├── mask_visib
        │   │   ├── depth
        │   │   ├── annotations.coco.json
        │   │   ├── scene_camera.json
        │   │   └── scene_info.json
        │   ├── 000002
        │   ├── 000003
        │   ├── 000004
        │   ├── 000005
        │   └── 000006
        └── task2
   

Description of Key Folders

Models

This directory contains the 3D models of objects used for synthetic data generation, rendering, simulation, and annotation. Each object is provided in multiple 3D file formats to support different workflows and software tools.

• 3d_scans: This folder contains the high-fidelity scanned 3D models of the physical objects used in the i-BADAS dataset.
• partial_scans: This folder contains top and bottom viewpoint-specific scans of the objects.

test

This directory contains the evaluation data for each object instance. It contains both anomaly-free test samples and test samples with defects or anomalous regions. It stores annotation information in COCO-style format and pose-related information for the test samples.

• good: This subfolder contains RGB images, depth data, and visible box object masks for normal, defect-free samples and includes scene-level metadata files which provide camera parameters and scene-specific information.
• anomaly: subfolder contains RGB images and depth data corresponding to anomalous samples. It additionally includes several mask folder like mask_all, mask_nuisance, mask_residual, and mask_visib. These masks provide different types of pixel-level annotations for the anomalous test images, including complete anomaly masks, nuisance-region masks, residual anomaly masks, and visible box object masks. The folder also contains scene metadata files.

  Figure 1: mask_all.

  Figure 2: mask_nuisance

  Figure 3: mask_visib (box)

  Figure 3: rgb (Anomalous)

  train

  This directory contains a complete real-world training scene captured using an RGB-D camera setup. The scene includes RGB images, depth images, visibility masks, object annotations, camera parameters, and scene metadata required for object detection, segmentation, pose estimation, and synthetic-to-real transfer experiments.

  • Task 1: This subfolder contains RGB images, depth data, and visible box object masks for normal, defect-free samples of empty_clean boxes for task 1 and includes scene-level metadata files which provide camera parameters and scene-specific information.
  • Task 2: This subfolder contains RGB images, depth data, and visible box object masks for normal, defect-free samples of empty_dirty boxes for task 2 and includes scene-level metadata files which provide camera parameters and scene-specific information.
  • rgb: Contains the RGB images captured for this scene.
  • mask_visib: Contains visible box object masks.
  • depth: Contains depth images corresponding to the RGB images.

    Figure 1: mask_visib (box).

    Figure 2: rgb (anomaly-free).

    License

    cc-by-nc-4.0

    Contact

    For questions, please contact the corresponding authors of the paper.