eduzhai > Applied Sciences > Engineering >

Iterate & Cluster Iterative Semi-Supervised Action Recognition

  • Save

... pages left unread,continue reading

Document pages: 11 pages

Abstract: We propose a novel system for active semi-supervised feature-based actionrecognition. Given time sequences of features tracked during movements oursystem clusters the sequences into actions. Our system is based onencoder-decoder unsupervised methods shown to perform clustering byself-organization of their latent representation through the auto-regressiontask. These methods were tested on human action recognition benchmarks andoutperformed non-feature based unsupervised methods and achieved comparableaccuracy to skeleton-based supervised methods. However, such methods rely onK-Nearest Neighbours (KNN) associating sequences to actions, and generalfeatures with no annotated data would correspond to approximate clusters whichcould be further enhanced. Our system proposes an iterative semi-supervisedmethod to address this challenge and to actively learn the association ofclusters and actions. The method utilizes latent space embedding and clusteringof the unsupervised encoder-decoder to guide the selection of sequences to beannotated in each iteration. Each iteration, the selection aims to enhanceaction recognition accuracy while choosing a small number of sequences forannotation. We test the approach on human skeleton-based action recognitionbenchmarks assuming that only annotations chosen by our method are availableand on mouse movements videos recorded in lab experiments. We show that oursystem can boost recognition performance with only a small percentage ofannotations. The system can be used as an interactive annotation tool to guidelabeling efforts for in the wild videos of various objects and actions toreach robust recognition.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×