FloodDET

Description

The FloodDET dataset is a large-scale, unified image collection designed to train deep learning models for the automatic detection and fine-grained quantification of urban flood levels. Addressing the disjointed annotation schemas of previous resources, this dataset provides a consistent, pixel-level re-annotation of 10,262 images exported in the standard MS COCO JSON format for seamless integration with major computer vision libraries. The dataset is split into training (7,477), validation (1,535), and test (1,250) sets, and incorporates non-flooded negative samples to minimize false positives. It employs a unique "regression-ready" 45-class schema that frames flood estimation as a fine-grained object detection task. By linearly mapping the 11 granular submersion levels (0 to 10) of four reference objects—Person (classes 1–11), Bicycle (12–22), Car (23–33), and Bus (34–44)—alongside a general "Flood" class (45), models can mathematically derive ground-truth regression targets directly from the class IDs. This harmonized dataset builds upon the extensive foundational data collection efforts of previous researchers. FloodDET consists of completely re-annotated raw imagery originally sourced from the DEEPFLOOD dataset (Chaudhary et al., 2020), FloodIMG (Karanjit et al., 2023), Mapillary Vistas (Neuhold et al., 2017), and MS COCO (Lin et al., 2014), which provided the essential non-flooded urban scene samples. We extend our full gratitude to these original authors for providing the raw imagery that made this unified resource possible. References: 1. Chaudhary, P., D’Aronco, S., Leitão, J., Schindler, K., Wegner, J., 2020. Water level prediction from social media images with a multi-task ranking approach. ISPRS Journal of Photogrammetry and Remote Sensing 167, 252–262 2. Karanjit, R., Pally, R., Samadi, S., 2023. Floodimg: Flood image database system. Data in Brief 48, 109164 3. Neuhold, G., Ollmann, T., Bulò, S.R., Kontschieder, P., 2017. The mapillary vistas dataset for semantic understanding of street scenes, in: 2017 IEEE International Conference on Computer Vision (ICCV), pp. 5000–5009. 4. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L., 2014. Microsoft coco: Common objects in context, in: European conference on computer vision, Springer. pp. 740–755.

Steps to reproduce

The dataset is exported in the standard MS COCO JSON format, making it directly compatible out-of-the-box with standard computer vision libraries such as pycocotools, Ultralytics (YOLO), MMDetection, and Detectron2.

Institutions

• Indian Institute of Technology Guwahati

  Assam, Guwahati

Categories

Licence