Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BugNIST -- a Large Volumetric Dataset for Object Detection under Domain Shift

Patrick Møller Jensen, Vedrana Andersen Dahl, Carsten Gundlach, Rebecca Engberg, Hans Martin Kjer, Anders Bjorholm Dahl

TL;DR

BugNIST addresses context shift in volumetric 3D object detection by training on isolated bug volumes and testing on densely packed mixtures where object appearance remains constant but surroundings differ. The dataset includes 9154 individual bug volumes across 12 classes and 388 mixtures, with center-point annotations used for testing, and baseline analyses are provided for context-shift detection. Three detection baselines (Detection U‑Net, 3D Faster R‑CNN, and nnDetection) are evaluated under three training strategies (single bugs, synthetic mixtures, crowded mixtures), using center-point matching with a bounding box inflated by a factor of $1.5$ for alignment. Results show that localization is achievable but class labeling degrades under context shift; U‑Net demonstrates robustness, while the other baselines struggle on mixtures, highlighting the need for context-aware methods. BugNIST thus offers a practical testbed for advancing 3D detection under context shift and supports extensions to segmentation, generative modeling, and morphology analysis in volumetric imaging.

Abstract

Domain shift significantly influences the performance of deep learning algorithms, particularly for object detection within volumetric 3D images. Annotated training data is essential for deep learning-based object detection. However, annotating densely packed objects is time-consuming and costly. Instead, we suggest training models on individually scanned objects, causing a domain shift between training and detection data. To address this challenge, we introduce the BugNIST dataset, comprising 9154 micro-CT volumes of 12 bug types and 388 volumes of tightly packed bug mixtures. This dataset is characterized by having objects with the same appearance in the source and target domains, which is uncommon for other benchmark datasets for domain shift. During training, individual bug volumes labeled by class are utilized, while testing employs mixtures with center point annotations and bug type labels. Together with the dataset, we provide a baseline detection analysis, with the aim of advancing the field of 3D object detection methods.

BugNIST -- a Large Volumetric Dataset for Object Detection under Domain Shift

TL;DR

BugNIST addresses context shift in volumetric 3D object detection by training on isolated bug volumes and testing on densely packed mixtures where object appearance remains constant but surroundings differ. The dataset includes 9154 individual bug volumes across 12 classes and 388 mixtures, with center-point annotations used for testing, and baseline analyses are provided for context-shift detection. Three detection baselines (Detection U‑Net, 3D Faster R‑CNN, and nnDetection) are evaluated under three training strategies (single bugs, synthetic mixtures, crowded mixtures), using center-point matching with a bounding box inflated by a factor of for alignment. Results show that localization is achievable but class labeling degrades under context shift; U‑Net demonstrates robustness, while the other baselines struggle on mixtures, highlighting the need for context-aware methods. BugNIST thus offers a practical testbed for advancing 3D detection under context shift and supports extensions to segmentation, generative modeling, and morphology analysis in volumetric imaging.

Abstract

Domain shift significantly influences the performance of deep learning algorithms, particularly for object detection within volumetric 3D images. Annotated training data is essential for deep learning-based object detection. However, annotating densely packed objects is time-consuming and costly. Instead, we suggest training models on individually scanned objects, causing a domain shift between training and detection data. To address this challenge, we introduce the BugNIST dataset, comprising 9154 micro-CT volumes of 12 bug types and 388 volumes of tightly packed bug mixtures. This dataset is characterized by having objects with the same appearance in the source and target domains, which is uncommon for other benchmark datasets for domain shift. During training, individual bug volumes labeled by class are utilized, while testing employs mixtures with center point annotations and bug type labels. Together with the dataset, we provide a baseline detection analysis, with the aim of advancing the field of 3D object detection methods.
Paper Structure (27 sections, 12 figures, 5 tables)

This paper contains 27 sections, 12 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Contributions
  3. Related work
  4. Data acquisition and processing
  5. Experiments
  6. Training strategies for detection under context shift
  7. Single bugs
  8. Synthetic mixtures
  9. Crowded synthetic mixtures
  10. Baseline detection models
  11. Detection U-Net
  12. Faster R-CNN
  13. nnDetection
  14. Evaluation
  15. Results
  16. ...and 12 more sections

Figures (12)

  • Figure 1: Overview of the two parts of the dataset. In (a), representative individual bug volumes are displayed with a class abbreviation, name, and number of volumes. In (b), a few volumes of bug mixtures are shown. Each bug is annotated with bug class and center point.
  • Figure 1: Volume renderings of the full scans and a single bug. The left shows the scanned bundle of tubes and the middle shows the same volume with higher transparency to reveal the bugs inside the tubes. The right shows a rendering of a single bug with low transparency to make even the air around the bug slightly opaque and highlights the crop-out cylindrical region containing a bug, air, and some cotton at the bottom. The region outside the crop-out cylinder is set to zero and is fully transparent.
  • Figure 2: Summary of volumes from scans of bug mixtures.
  • Figure 2: Overview of the downscaled BugNIST individual bugs with sizes x512, x256, x128, and x64. Images are maximum projections from the side (first row of each block) and from the top (second row of each block). These images were used for classification experiments also included in this supplementary material.
  • Figure 3: The top line shows the automated annotation pipeline for individual bugs. By thresholding, bounding boxes and segmentation masks are easily obtained. The bottom shows the synthetic mixtures used for the more advanced training strategy.
  • ...and 7 more figures