Few-Shot Object Detection for Classifying Reported Urban Problems

Authors:

Abstract
These days, one of the government-citizen engagements is a digital report submitted by citizens. Urban problem detection is a process to classify images reported in an application to be followed up by the Government. The detection is crucial for maintaining livable and sustainable cities. However, annotating large-scale datasets for object detection models is labor-intensive. This paper proposes a few-shot object detection model for urban problem detection, overcoming data limitations. Our model demonstrates promising performance across 13 selected categories, including 10 base classes and 3 novel classes, using a dataset from the Jakarta Kini (JAKI) application. We evaluate the model on familiar (base classes) and unseen (novel classes) categories, showcasing its generalization capabilities. With increased instances (n-shot), our model achieves high average precision (AP) scores, with the highest performance observed at 30-shot. The validation evaluation yields AP scores of 54.5% for all classes, 65.3% for base classes, and 71.6% for novel classes. Similarly, the test evaluation results in AP scores of 42.5% for all classes, 51.4% for base classes, and 69.8% for novel classes. The performance differences between the validation and test datasets stem from unique features and category challenges. Our few-shot object detection model holds promise for real-world urban problem detection and can be extended to other urban-related tasks and datasets.