| |||||||||
CVPRW 2017 : The 1st Workshop of Open Domain Action Recognition in conjuction with CVPR 2017 | |||||||||
Link: http://sesame.comp.nus.edu.sg/workshop/odar2017/ | |||||||||
| |||||||||
Call For Papers | |||||||||
Action recognition has been an active computer vision and machine learning problem in the last two decades. The research works in this area mostly focus on laboratory captured action datasets, which exhibit visually similar and controlled environment, and have finite set of action classes and actors. In recent years, due to the availability of big data and egocentric devices, many research efforts are directed towards designing algorithms to handle such data, which has large variation in action classes, environment, captured conditions and so on. For the laboratory captured action datasets, despite the literature having reported close to perfect classification accuracy, the successful adoption of state-of-the-art action recognition algorithms in real-world use case scenarios, such as video surveillance, is still far from prevalent. One of the key bottleneck in the conventional approach is that its ignorance of the open domain constraint (i.e., the deployment environment are likely not to be seen in the training data space).
Challenge Track This workshop will organize an open domain action recognition challenge to address this issue. Based on existing open source constrained action recognition datasets, we have built a new dataset to address this challenge. The dataset consists of multiple publicly available datasets with carefully selected action classes that are common across these datasets. In this challenge, we have designed a protocol to ensure the training set, validation set, and test set contains action samples extracted from different domain (camera view or database). Therefore, we guarantee that the data in the validation set and test set have never appeared in training data. Overall, we have selected 11 common action classes and an additional class for all other action samples (i.e., "unknown" class). This workshop shall be an opportunity for researchers from related fields to revisit the action recognition problem under this new perspective. Specifically, given the carefully captured dataset, how do we design algorithms to perform well under target domain, which are not presented during the training stage. It will drive the research to focus on the open domain constraint, as well as to drive the practical usage of action recognition algorithm in real-world surveillance or monitoring applications. To access the challenge dataset, please Register for the access information. Regular Track The regular track is open for paper submission on topics related to open domain classification problems. The authors are welcomed to submit original and innovative papers on the related areas. We aim for broad scope, topics of interest include but are not limited to: Representation and feature descriptors Detection and Tracking of interest points Modelling of human motion Open domain learning and classification Cross-view learning and classification Open-set learning and classification One-shot learning and classification Zero-shot classification Detection of actions and activities in videos Interaction between humans and objects Deep learning for open-domain action recognition The deadlines for this CFP are the following: Regular paper submission: 2017 March 31 Notification of acceptance: 2017 April 20 Camera-ready submission: 2017 April 27 Challenge test set release: 2017 May 1 Challenge submission: 2017 June 1 Best Regards, ODAR 2017 Organizing committee |
|