Towards fine-grained object-level damage assessment during disasters

Social media can play an important role in current-day disaster management. Images shared from the areas may include objects relevant to operations. If these are identified correctly, they offer a preliminary damage assessment report and situational awareness for response recovery. This research is carried out collaboration with Community Emergency Response Team (CERT) understand state-of-the-art object detection model’s capability detect multi-hazard scenes posted on social media. Specifically, 946 images were collected during major earthquake hurricane disasters. All inspected by trained volunteers CERT and, 4,843 analyzed applicability specific functions The feedback provided helped determine existing key strengths weaknesses led development of taxonomy support functions. Lastly, using subset classes taxonomy, instance segmentation dataset developed fine-tune models detection. Empirical analysis demonstrates promising applications transfer learning