A stable and efficient dynamic ensemble method for pothole detection

Abstract

Roads can develop potholes over time, posing hazards to traffic. However, regular road damage inspections is challenging due to the high cost of road surveys. By applying object detection models on footage acquired from dashboard cameras installed in garbage trucks that operate across the city, we can conduct road surveys at a low cost. In our previous work we introduced the Ensemble of Classification Mechanisms (ECM), which suppresses false positives by cross-verifying objects detected by an object detection model using a different image classification model. However, ECM faces challenges in achieving both fast inference speed and high detection performance simultaneously. It also struggles in environments where roads vary in their suitability for false positive suppression. To address these issues, we propose the Dynamic Ensemble of Classification Mechanisms (DynamicECM). This approach utilizes ECM selectively, enabling stable inference with minimal false positive suppression. To evaluate our new method, we constructed an evaluation dataset comprising objects that cause false positives in pothole detection. Our experiments demonstrate that ECM achieves higher precision, average precision (AP), and F1 scores compared to existing methods. Furthermore, DynamicECM improves the trade-off between speed and detection performance, outperforming ECM, and achieves stable inference even in challenging datasets where ECM would falter. Our method is highly scalable and expected to contribute to the stability and efficiency of inference across various object detection models. In our previous work we developed an Ensemble of Classification Mechanisms (ECM), which suppresses false positives by rechecking objects detected by an object detector with a different image classification model. However, ECM cannot achieve both fast inference speed and high detection performance at the same time. It also struggles in environments that have a mixture of roads suitable for false positive suppression and unsuited for false positive suppression. To solve these problems, we propose “Dynamic Ensemble of Classification Mechanisms”. Since this method uses ECM only when deemed necessary, stable inference can be achieved efficiently without excessive suppression of false positives. In order to evaluate our new method, we constructed an evaluation dataset that includes objects that cause false positives in pothole detection. Our evaluation experiments show that ECM achieves higher precision, AP, and F1 compared to existing methods. In addition, DynamicECM improves the trade-off between speed and detection performance better than ECM, and achieves stable inference on datasets that would ECM would struggle on. Our method is highly scalable and expected to contribute to the stability and efficiency of inference for various object detection models.