Damage Building Detection from Crowdsourcing Images Using Transfer Learning
In this presentation, we used images from public domain, i.e. Twitter, Getty images etc to train a deep learning model to recoganize damaged buildings caused by earthquakes. The goal of building this model is to label damaged buildings from crowdsourcing images fast after an earthquake, and combining with geolocation information, this can be used to verify the damage in a city. We started from VGG19 model, and use transfer learning to fine tune the last block CNN layers. We tested the model both in real time and offline on Twitter, and the model achieves really good performance in terms of precision and recall. In addition, we also try to visualize what the model learns to make the decision via Grad-CAM. We will also discuss some challenges we found, such as night scenes, arial images etc.
Presenting Author: Qingkai Kong
Student Presenter: No
Day: 4/19/2021
Time: 3:45 PM - 4:45 PM Pacific
Additional Authors
Gaurav Chachra chachra@berkeley.edu University of California, Berkeley |
Qingkai Kong Presenting Author Corresponding Author kongqk@berkeley.edu Berkeley Seismological Laboratory |
Jim Huang ccjimhuang@gmail.com AT&T |
Srujay Korlakunta srujay@berkeley.edu University of California, Berkeley |
Alexander Robson a.robson@berkeley.edu University of California, Berkeley |
Jennifer Grannen jenngrannen@berkeley.edu University of California, Berkeley |
Richard Allen rallen@berkeley.edu Berkeley Seismological Laboratory |
|
|
Damage Building Detection from Crowdsourcing Images Using Transfer Learning
Category
When Seismology Meets Machine Learning, Data Science, HPC, Cloud Computing and Beyond
Description