Abstract
When deploying a model for object detection, a confidence score threshold is chosen to filter out false positives and ensure that a predicted bounding box has a certain minimum score. To achieve state-of-the-art performance on benchmark datasets, most neural networks use a rather low threshold as a high number of false positives is not penalized by standard evaluation metrics. However, in scenarios of Artificial Intelligence (AI) applications that require high confidence scores (e.g., due to legal requirements or consequences of incorrect detections are severe) or a certain level of model robustness is required, it is unclear which base model to use since they were mainly optimized for benchmark scores. In this paper, we propose a method to find the optimum performance point of a model as a basis for fairer comparison and deeper insights into the trade-offs caused by selecting a confidence score threshold.
Highlights
Object detection is still a hot topic in academia and industry
Instance segmentation is generally preferred in high-aspect-ratio object detection, which is often prevalent in these applications; in some circumstances, where computational power requirements are limited or there is a lack of datasets classical object detection must do be sufficient enough
The selection of evaluation metrics (Section 2.3) requires that a dataset provides data suitable for instance segmentation and not just bounding boxes as this is required by the Probabilitybased Detection Quality (PDQ)
Summary
Object detection is still a hot topic in academia and industry. The term “object detection” yields ca. 33,400 search results on Google Scholar (excluding patents and citations) and arXiv.org lists 410 papers uploaded (title and abstract searched) for 2020. Sensor development has led the way to ground structures detection [3] and analysis from remote sensing satellite imagery data One example of this is bridge dislocation analysis, where a high degree of displacement detection precision is achieved [4,5]. This gives promise for remote sensing maintenance analysis for power grid and oil/gas pipelines as well. These applications have minimum requirements with respect to model robustness while maintaining deployability on edge devices. Instance segmentation is generally preferred in high-aspect-ratio object detection, which is often prevalent in these applications; in some circumstances, where computational power requirements are limited (e.g., on-board processing) or there is a lack of datasets classical object detection must do be sufficient enough
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have