Core Concepts
Domain shift challenges in object detection addressed by BugNIST dataset.
Abstract
The BugNIST dataset introduces a solution to domain shift challenges in object detection by providing annotated training data for deep learning algorithms. The dataset comprises individual bug volumes and mixtures, aiming to advance 3D object detection methods. Training models on isolated objects and testing on mixtures reveal the impact of context shift on performance.
Directory:
Introduction
Motivation for labeled data in volumetric imaging.
Proposal to address object detection and classification.
Data Extraction and Processing
Creation of BugNIST dataset with individual bugs and mixtures.
Importance of context shift in domain adaptation.
Related Work
Comparison with existing datasets for domain shift studies.
Experiments and Evaluation
Baseline models: U-Net, Faster R-CNN, nnDetection.
Training strategies: single bugs, synthetic mixtures, crowded mixtures.
Results Analysis
Performance evaluation on real mixed scans.
Discussion and Conclusion
Impact of context shift on model performance.
Potential applications beyond object detection.
Stats
"This dataset is characterized by having objects with the same appearance in the source and target domain."
"BugNIST contains 9154 micro-CT volumes with individual bugs and 388 volumes containing a mixture of several bugs."
"In total, 9573 volumes make up the BugNIST dataset."
Quotes
"The term context shift refers to a shift in the data distribution between training and inference."
"Solutions will enable an automated annotation of data for volumetric object detection."