Object detection from images plays an important role in ITS applications, such as autonomous-driving systems. A concern for recent object detection models based on neural networks is the black box nature of the neural networks, which should be addressed in mission critical ITS applications. In this paper, we propose a framework of partial image retrieval to analyze and debug a target detection model. Given an image with incorrectly detected objects, our method discovers similar images that also contain incorrectly detected objects from a large unlabeled image database, which is relatively easier to collect than annotating all images. These similar images provide us with insights into the misdetection by object detection models, which will be useful in resolving the issue. We qualitatively and quantitatively demonstrate that the proposed framework effectively outputs similar doubtful images using real-world datasets. Specifically, we show that, by querying an image with a false positive detection, we can efficiently find images with false positives from a large unlabeled image database using the proposed framework.
Finding Misdetections in Unlabeled Image Database: Toward Reliable Object Detection
2021-09-19
2749645 byte
Conference paper
Electronic Resource
English
REPRESENTATION LEARNING FOR OBJECT DETECTION FROM UNLABELED POINT CLOUD SEQUENCES
European Patent Office | 2024
|MINING UNLABELED IMAGES WITH VISION AND LANGUAGE MODELS FOR IMPROVING OBJECT DETECTION
European Patent Office | 2023
|Finding where the earth is reliable
Online Contents | 2012
British Library Online Contents | 2009
|