Evaluation of Image Processing Techniques on Outdoor Photogrammetry Data

Sammanfattning

This thesis investigates and evaluates three image processing techniques for improving photogrammetry. This is done with an outdoor dataset from the Tanks and Temples benchmark. The study focuses on geometric accuracy, i.e. how close the reconstruction is to the ground truth, and not on aesthetically pleasing, but incorrect, results. The three techniques used are deep learning models, addressing one common problem each. Ambient lighting normalization is attempted with IFBlend to mitigate uneven scene illumination. Background removal is done through U2-Net to reduce processing times. Upscaling of images is achieved through super resolution with REAL-ESRGAN in an attempt to extract more model features. To evaluate these techniques, a pipeline was developed, consisting of an image processing step and then a 3D reconstruction step with the photogrammetry software Meshroom. After reconstruction, the model is analysed using the software CloudCompare for geometric accuracy, and logging scripts are used to measure computational resources. In addition, the auto-generated logs from Meshroom are analysed to find a deeper explanation of the reconstruction results. With the main results, I propose a reproducible pipeline for applying and analytically evaluating image processing techniques on photogrammetry data. The evaluation results show that background removal reduced computation time by 35% while halving computational use compared to the baseline of no image processing applied. It has comparable geometric accuracy in areas not affected by segmentation errors, but runs the risk of accidentally removing parts of the subject. Ambient lighting normalization did not improve in any metric, most likely due to the pre-trained model not generalizing to outdoor conditions. Super resolution improved geometric accuracy with an RMS of 4.69 mm and outlier ratio of 0.3% compared to the baseline with 8.71 mm and 2.1%. Though, this came at a significant computational cost and with a computation time primarily limited by the hardware at 14.5 hours. Overall, super resolution and background removal show promise to be included in drone-based photogrammetry pipelines, while ambient lighting normalization would require further development before practical application. The deeper explanation derived from the logs supports that the weak spots in the performances stem from the inconsistent lighting normalization for ambient lighting normalization, misclassification for background removal, and RAM hardware limitations for super resolution.