Active Learning for 3D Mesh Semantic Segmentation
summary
Claims that Monte‑Carlo Dropout applied to PointNet++ meaningfully reduces annotation needs for 3D mesh/pointcloud semantic segmentation and shows where it helps: scene, segment, and point selection. Experiments on the scanner data set report MC Dropout beating random selection after ~90 scenes, needing roughly about 40% of segments to approach full‑data performance (around 0.45 mIoU), and require 20× inference runs plus costly retraining — the method works but is bottlenecked by inference and retrain speed.
AI-written summary · about
chapters (15)
- 0:00Active learning for 3D mesh segmentation
- 0:31Rising interest in 3D data
- 3:023D semantic segmentation tasks
- 4:32Point-based methods in 3D mesh
- 6:44Sparse convolution for 3D data
- 8:11PointNet++ overview
- 12:11Sampling and set abstraction layers
- 16:31Active learning workflow
- 19:28Monte Carlo Dropout for uncertainty
- 21:16Experiments on ScanNet
- 24:30Scene-level active learning results
- 26:03Segment-level active learning results
- 27:50Point-level active learning results
- 28:55Future work and bottlenecks
- 31:29Q&A
related talks
42mDeep Learning approaches meet 3D data
39mBayesian Deep Learning with Edward (and a trick using Dropout)
47mThe whole truth about using deep learning in medical imaging
51mEfficient Deep Learning with Humans in the Loop
35mAdjusting 2D prediction models to be usable for 3D objects
27mUnderstanding Deep Neural Networks
58mDeep Learning in Real Life
24mAutomating Porosity Detection with Deep Learning