Abstract
M.Ing. (Electrical Engineering)
Dense stereo photogrammetry techniques are presented as a solution to 3D scene reconstruction, providing depth measurements for many application domains where passive images can be taken. Often stereo algorithms evaluate their reconstructions against a single dataset to inform performance characteristics, calculated as a function of the accuracy across different characteristic regions.
In this dissertation, the generalisability of dense reconstructions is evaluated across different scene environments, highlighting the role of both accuracy and density in formulating performance metrics. These combined metrics identify possible bias, whereby one metric may be inflated at the cost of the other. Experimental tests are conducted on six state-of-the-art algorithms in the presence of varying scene conditions and datasets of different ground truth formations, offering a diverse representation of approaches that follow the dense stereo taxonomy. The top-performing algorithm is further experimented on, by applying various pre-processing and post-processing techniques to identify conditions that reduce disparity map estimation errors.
Our findings identify that ELAS, a simple Bayesian inference model, offers the most value as it provides highly accurate disparity map reconstructions that are semi-dense. Alternatively, the local aggregation based Cost-Volume filter offers a unique density completeness characteristic, with competitively high accuracies. ELAS pre-processing identifies a simple noise removal to offer the best accuracy improvement, while density gains can be achieved with illumination and edge enhancement techniques. Post-processing via consistency cross-checking achieves small accuracy gains, while interpolation techniques offer large density improvements at a small cost to the disparity map accuracy.