- Utilized web scraping skill to collect metadata from available databases. 4,000+ images of 1,228 buildings were obtained through Google Street View API.- Implemented faster RCNN-Inception/RCNN-Resnet to train an object detection model that can compute the number of floors.- Labeled 1,000+ images with bounding boxes and treated them as the training set.- Computed the real/pixel length ratios through OpenCV for inferencing floor heights of 1,228 buildings.- Evaluated… Show more - Utilized web scraping skill to collect metadata from available databases. 4,000+ images of 1,228 buildings were obtained through Google Street View API.- Implemented faster RCNN-Inception/RCNN-Resnet to train an object detection model that can compute the number of floors.- Labeled 1,000+ images with bounding boxes and treated them as the training set.- Computed the real/pixel length ratios through OpenCV for inferencing floor heights of 1,228 buildings.- Evaluated earthquake-safety of 1,228 buildings in Los Angeles through high-performance computing under Linux system.- Created an excel database with 1,000,000+ buildings’ information and simulation results for decision makers (e.g., government, insurance company, and property owner). Show less

2019 AUVSI Competition (UAS@UCLA)- Ranked 22/68 in overall tasks.- Created an object detection model of faster RCNN in the unmanned aerial system to recognize colored alphanumeric character painted on a color shape.- Labeled 1,400+ Google Maps images with bounding boxes and treated them as the training set.

Peer Hub ImageNet Challenge -Applied Transfer Learning with the deep convolutional neural network (e.g., VGG16, VGG19, Inception, and ResNet) to train an image classification model for structural damage level.-Utilized Keras/TensorFlow to develop the classification model and use Google Cloud Platform for the training process. 17,000+ images were treated as the training set and 8 prediction tasks with at most 4 categories have been performed.-Implemented data argumentation and merged… Show more Peer Hub ImageNet Challenge -Applied Transfer Learning with the deep convolutional neural network (e.g., VGG16, VGG19, Inception, and ResNet) to train an image classification model for structural damage level.-Utilized Keras/TensorFlow to develop the classification model and use Google Cloud Platform for the training process. 17,000+ images were treated as the training set and 8 prediction tasks with at most 4 categories have been performed.-Implemented data argumentation and merged 5 state-of-the-art neural networks to achieve 90% accuracy for classification with 3 categories. Show less

- Applied PointNet for classifying seismic vulnerable buildings using city-scale point-cloud datasets.- Labeled 1,000,000,000+ point-cloud data in Santa Monica for training, validation and testing.- Performed sensitivity analysis and achieved 90%+ accuracy and the intersection over union.- Combined point clouds and GIS to obtain address of predicted seismic vulnerable buildings.- Implemented opensource iModel.JS for 3D visualization.