Browse Teach Curate

Masked Autoencoding on Event-Camera Point Streams for Action Recognition

Treats an event-camera stream as a 3D point cloud in (x, y, time) and pre-trains a transformer encoder-decoder by masking and reconstructing event patches, learning representations that transfer to action and gesture recognition. Introduces an inlier plane-fitting model that selects noise-free patch centers, replacing farthest-point sampling, with PointNet patch embeddings and a Chamfer-distance reconstruction loss.

Data Text

Event Masked Autoencoder: Point-wise Action Recognition with Event-Based Cameras Jingkai Sun1,2,∗

Event cameras (dynamic vision sensors) output asynchronous streams of points, each carrying a pixel location, a timestamp, and a polarity indicating whether brightness rose or fell, rather than frame…