Object Detection via Point Cloud and Deep Learning

Building and Infrastructure Defect Detection and Visualization Using Drone and Deep Learning Technologies

full-text on https://doi.org/10.1061/(ASCE)CF.1943-5509.0001652

Window Detection Ver. 2

Dr. Yuhan Jiang 2020/11/04

Highlights:

The developed application (running in background) automatically captures screenshots and processes them

The manual participation was crafting few labels

The point cloud was generated from Autodesk ReCap Pro

The pixelwise segmentation was powered by U-net

The U-net model training only used few orthographic (front, back, left, and right) view images

The data augmentation included rotation, resize, perspective, and flip

The “contours” function was powered by OpenCV

Testing Result

Captured image Pixelwise segmentation

Training Data Set

Image Label