Abstract: Object Detection of Damaged Buildings in disastrous event is important for aiding and reconstruction. Current approaches for buildings damage assessment include ... / However, these pre-trained model is still lack of generalized capability which fails to timeliness of detection. / In this work, we propose a generative model ModelName to manufacture post-disaster image at global scale, be it GBD. / We find that after further training SOTA models on GBD, the performance of models show great improvements.

Backbone Network: Unet(Learning Knowledge) + Diffusion(Generation)

Zhang, Y., Shi, X., Li, D., Wang, X., Wang, J., & Li, H. (2023). A Unified Conditional Framework for Diffusion-based Image Restoration. 37th Conference on Neural Information Processing Systems (NeurIPS 2023)

Shortage of labeled building dataset at global scale.

The feature of remote sensing images on buildings varies greatly from here to there.

Pre-trained models strongly rely on domain-specific dataset for further training/supervised fine-tuning to improve the performance for specific downstream tasks.

In terms of Catastrophe Detection from Remote Sensing Images, preview work focus on (Li et al., 2024)

• Building Type Classification
• Building Change Detection

Li, Q., Mou, L., Sun, Y., Hua, Y., Shi, Y., & Zhu, X. X. (2024). A Review of Building Extraction From Remote Sensing Imagery: Geometrical Structures and Semantic Attributes. IEEE Transactions on Geoscience and Remote Sensing, 62, 1–15. https://doi.org/10.1109/TGRS.2024.3369723

We propose a generative model to synthesize labeled building damage dataset.

SOTA models further train on our dataset shows considerable improvement in zero-shot for downstream tasks like segmentation, change detection and scene classification, especially in regions that has never been seen in the training dataset.

• Incorporated Language Model/Contextual Cues
• A generalized model through Mixture of Experts/Modality Bridging