Monocular Depth Estimation For Satellite Depth Map Enhancement

Sammanfattning

Monocular Depth Estimation (MDE), the task of estimating depth from a single image, is a prominent topic in computer vision. However, no prior work has explored using MDE to enhance satellite 3D depth maps. This thesis investigates the application of MDE within the domain of Remote Sensing to predict depth maps with sharp edges and fine details. Several models, fine-tuned from Depth Anything V2, are trained with various loss functions on both real and synthetic satellite imagery from San Antonio and San Francisco. The corresponding ground truth is rendered from high-resolution aerial 3D models. These models are then validated and tested on comparable datasets from Las Vegas and Linköping. Additionally, this study explores how depth maps derived from stereo pairs can be improved by combining them with predicted depth. These are combined with three primary methods: Mean, Frequency, and Registration Alignment. This thesis concludes that the top-performing fine-tuned models significantly outperform the general-purpose Depth Anything V2 model. The results also indicate that pretraining on synthetic data can improve generalization. Moreover, incorporating a gradient-based loss term has proven beneficial for model training. Another key observation was that zoomed-in inputs led to predictions with finer detail. The best predictions successfully improved the baseline depth maps, generated from stereo pairs, across all three combination strategies. For all methods, minimum height correction was beneficial. Among these, the frequency alignment method achieved the best overall performance. In summary, the enhanced depth maps exhibited sharper edges and a closer resemblance to ground truth data. These findings suggest that MDE holds strong potential for improving 3D model quality in remote sensing applications.