Multi-view damage inspection using single-view damage projection

Single-view computer vision models for vehicle damage inspection often suffer from strong light reflections. To resolve this, multiple images under various viewpoints can be used. However, multiple views increase the complexity as multi-view training data, specialized models, and damage re-identification over different views are required. In addition, traditional point cloud applications require large computational power, being impractical for edge computing. Therefore, multi-view damage inspection has not yet found its way into practical applications. We present a novel approach that projects the results from widely available single-view computer vision models onto 3D representations, to combine the detections from various viewpoints. With this, we leverage all advantages of multi-view damage inspection, without the need for multi-view training data and specialized models or hardware. We conduct a practical evaluation using a drive-through camera setup, to show the applicability of the methods in practice. We show that our proposed method successfully combines similar damages across viewpoints, reducing the number of duplicate damages by almost 99%. In addition, we show that our approach reduces the number of false positives by 96%. The proposed method leverages the existing single-view training data and single-view deep learning models to make multi-view inspection more accessible for practical implementations.