The preparation of a large and diverse dataset is essential for training a robust deep learning model. However, there are instances where certain data are theoretically possible but challenging to observe in reality (e.g., traffic light violations). We refer to these unique instances as 'Scarce Realities', highlighting their rarity and the difficulties they present in data collection and model training. One effective and emerging approach involves using generative models to generate and augment such data. In this study, we demonstrate the promising potential of combining object detection models with simple image generation models as a way to generate fake videos. We achieve this by partially editing existing videos to artificially create 'Scarce Realities', using the generation of fake dashboard camera footage of traffic light violations as an example.

Road damage such as potholes may occur on roads due to aging, which may affect traffic. Periodic inspections of road damages are difficult due to the high cost of road surveys. The development of a system that automatically detects potholes and other road damages from dash cam images can allow inexpensive road inspections and can overall improve the problem of the long-term overlook of road damages. Last year, we conducted a demonstration experiment in Edogawa City, Tokyo, using an existing image-based road damage detection method. From that experiment, we found that the detection of potholes on actual roads often causes false positives in detecting shadows and manholes. In this study, we propose a method to reduce false positives in pothole detection, which was considered to be a problem through the demonstration experiment. Since the evaluation based on a pothole-only dataset is not practical, we constructed a dataset for evaluation by adding shadow and manhole images. Our method consists of two main components: data augmentation and an ensemble of classification mechanisms for object detection models. The result of the test on the reconstructed pothole dataset showed that the Average Precision (AP), which is a measure to evaluate the performance of object detection, and F1, which is the harmonic mean of precision and recall, were improved compared to the existing method. Our new method is expected to be an effective pipeline for tasks and situations where false positives are likely to occur and where false positives are more considered as an issue than false negatives, given that they are not dependent on the domain of potholes.