Use static classifiers for dynamic point cloud tasks (3D) and use action classifiers for temporal anomaly detection (2D)

​This lecture will cover two papers by the author in the fields of motion understanding in 3D and video.

​(Paper 1) Can we painlessly modify the classifier of static point clouds to recognize a dynamic sequence of point clouds? To separate 3D motions without explicitly tracking correspondences, we propose a kinematic inspired neural network (Kinet) by generalizing the kinematic concept of ST-surfaces to the feature space.

​(Paper2) Can we train a fully-supervised action classifier to detect video abnormalities in a weakly-supervised manner? From the perspective of learning with noisy labels, we propose a graph convolutional label noise cleaner and propagate supervisory signals from high-confidence snippets to low-confidence ones.

​
The presentation is based on the speaker's paper and project:

• ​No Pain, Big Gain: Classify Dynamic Point Cloud Sequences with Static Models by Fitting Feature-level Space-time Surfaces (CVPR 2022)
  https://github.com/jx-zhong-for-academic-purpose/Kinet
  https://arxiv.org/abs/2203.11113

• ​Graph Convolutional Label Noise Cleaner: Train a Plug-and-play Action Classifier for Anomaly Detection (CVPR 2019)
  https://github.com/jx-zhong-for-academic-purpose/GCN-Anomaly-Detection
  https://arxiv.org/abs/1903.07256

​Presenter Bio:

​Jiaxing Zhong is a Ph.D. student in Computer Science at the University of Oxford, with research interests in machine learning and computer vision (e.g., 3D vision). He holds a Master's degree in Computer Science from Peking University in 2020.

​------------
Find us at:

• ​All lectures are uploaded to our Youtube channel

• ​Newsletter for updates about more events

• ​Sub-reddit for discussions

• ​Discord server for, well, discord

• ​Blog https://2d3d.ai