charlesq34 / pointnet

PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
Other
4.76k stars 1.45k forks source link

PointNet Classification Failure Causes and Solutions #184

Open raekyu opened 5 years ago

raekyu commented 5 years ago

Dear Charles:

I am a master student at Seoul National University of Science and Technology in Seoul, Korea. My lab is using PointNet to classify point clouds for architectural elements such as doors, windows, walls, slabs, etc. We have had success in distinguishing different subtypes of doors (e.g., single door, double door, sliding door, etc.)

However, we are having some difficulty in classifying wall elements. We want to distinguish walls as generic walls, wall with door openings and window openings. Sample elements are shown in below. For some reason PointNet cannot seem to distinguish them, and we are assuming that it cannot discern the ‘openings’ (holes) in the walls. Our classification results show that most are misclassified as generic walls.
Could you provide some advice on what the cause may be? Your input would be greatly appreciated and help us in moving forward with our research.

Thank you, Raekyu Jung MS Student Civil Engineering Department Seoul National University of Science and Technology Email : forb1214@naver.com

*IFC image

-General walls image

-Walls with window openings image

-Walls with door openings image