Estimating Absolute Flood Depth from Submerged Vehicles Using Image Retrieval Method Guangan Chen, Michiel Vlaminck, Brian Booth, Anh Minh Truong, Michiel De Baets, Wilfried Philips, Hiep Luong Image Processing and Interpretation (IPI), IMEC research group at Ghent University, Belgium Email:
[email protected] Floods cause major losses by damaging infrastructure and endangering lives, making them one of the most costly natural disasters worldwide. Traditional monitoring with sensors or satellites is expensive and limited in resolution, while abundant street-level images with vehicles as depth references offer a scalable alternative. However, most existing methods provide only discrete level estimation, and no dataset with ground truth flood depth is available for evaluation. In light of this, this study aims to estimate flood depth at centimeter-level accuracy based on submerged vehicles. Motivation •We propose a novel flood depth estimation pipeline, where partially submerged vehicles are first segmented using a hybrid approach that integrates Mask R-CNN and the Segment Anything Model (SAM). Flood depth is then estimated through image retrieval from an image retrieval dataset. •We propose a viewpoint sampling strategy to create a large-scale image retrieval dataset of partially visible vehicles for use in image retrieval-based flood depth estimation. The dataset comprises 6,912 images rendered from diverse viewpoints and across varying flood depth. Method Examples of segmented images of flooded vehicles from our image retrieval dataset. The flood levels from left to right are 1, 2, 5, and 8. Illustration of the discretization of flood depth into ten uniform intervals relative to the height of the reference car (a white 2022 Peugeot 308) in the dataset, and the corresponding nine flood levels used for evaluation. Illustration of the camera distribution used for rendering the 3D car model in the image retrieval dataset. (a) Side view; (b) Top view. To ensure diverse viewing angles for the rendered images, cameras are sampled across the surface of an upper hemisphere using a layered scheme, where each elevation layer is assigned a fixed number of cameras with azimuth angles shifted relative to adjacent layers. For a better visualization in this figure, the number of elevation layers and cameras per layer are both set to 10, while they are 10 and 72 for creating the dataset, respectively. (a) (b) Evaluation •Synthetic dataset of 200 images: To generate diverse and realistic flooded scenes with ground truth flood depth, we leveraged the inpainting capability of the large multimodal model, Stable Diffusion model, implemented via the Hugging Face Diffusers library. •Real-world dataset of 200 images: We developed a two-step GUI for annotating flood depth in 3D space and conducted a subjective study to collect annotation data. The GUI interfaces for collecting the real-world dataset. The progress of the subjective study is shown at the top, with the annotated image on the left and the operation panel on the right. (a)Step 1: users are asked to perceptually select the reference vehicle that best matches the vehicle closest to the camera and appearing largest in the image, based on visible components and overall shape of the target vehicle. (b)Step 2: Users adjust the height of the flood volume in 3D space by moving the slider handle or using keyboard inputs until the waterline in the 3D view perceptually aligns with the waterline observed in the reference 2D image. (a) (b) Results Conclusion We applied stable diffusion model to create a synthetic dataset of submerged vehicles with ground truth centimeter-level flood depth. In addition, we designed and developed a GUI for annotating a real-world dataset in 3D space. We proposed a retrieval-based framework for estimating flood depth from submerged vehicle images, achieving lower error and greater robustness than existing methods on both synthetic and real-world datasets. Version 2
