Towards improving synthetic-to-real image correlation for instance recognition in structure monitoring

Data-driven approaches in deep learning are inevitable in instance recognition applications in the absence of object description models. To alleviate the cost associated with the gathering of the tremendous amount of data needed to train reliable algorithms, the potential use of synthetic data has been envisioned. The latter solution, however presents a major setback labeled as "domain gap" in the field of computer vision. It refers to the ability of detection models to recognize the artificial nature of photo-realistic rendering which poses a serious issue for learning on synthetic images alone. This work takes on a parametric approach in order to assess the influence of common techniques in image processing on the effectiveness of object recognition. To illustrate, their effect the detection of defects in pipeline infrastructures is considered as a study case. Training and test datasets are altered using filters and color processing techniques and detection effectiveness is measured in each configuration. Recommendations are drawn from the results for the future efficient use of synthetic data in recognition model training.