Login Form

Editors

Activity recognition results on UCF Sports and Holywood2

Table above shows the results, obtained on UCF Sports dataset (http://crcv.ucf.edu/data/UCF_Sports_Action.php). We report recognition rate with respect to the number...


Read More...

Computational efficiency and parallel implementation

The developed algorithms are computationally effective and the compositional processing pipeline is well-suited for implementation on massively parallel architectures. Many...


Read More...

Motion hierarchy structure

Our model is comprised of three processing stages, as shown in the Figure. The task of the lowest stage (layers...


Read More...

Server crash

After experiencing a total server failure, we are back online. We apologize for the inconvenience - we are still in...


Read More...

L1: motion features

Layer L1 provides an input to the compositional hierarchy. Motion, obtained in L0 is encoded using a small dictionary.


Read More...
01234

Task 2.2: Performance evaluation of the motion model

Regardless of their nature (spatial, temporal, or spatio­temporal), low­level descriptors will depend on a number of parameters, which will determine various properties of the model, such as a spatial and temporal reach of low­level descriptors, and will influence the overall properties of the hierarchical compositional model. It is possible that other levels of hierarchical model will be able to compensate for inadequately tuned low­level descriptors, however, we will incorporate the mechanisms to evaluate and alleviate such situations. In related research, [Pinto2009] has shown that in shape­based compositional models with huge parameter space, there exist “rare­events” ­­ small regions in the parameter space that exhibit significantly better performance than majority of combinations of parameter values. We expect that identification of these rare­event regions will play a crucial role in the performance of our method. In this task, we will need to devise the criteria, which will be used in adjusting the properties of low­level descriptors. This criteria should be independent from the rest of the hierarchy. It is expected that such criteria will be based on the information conveyed by the obtained descriptors, for example how well are they able to distinguish between random variations (noise, surface motion, object motion) and human motion. We expect that there exists a set of motion primitives, which are more complex than simple spatial, temporal or spatio­temporal motion descriptors, but could be nevertheless universally applied to any action or activity recognition task, regardless of its nature.

We expect that the introduction of time into the hierarchical compositional model could be done in two separate ways. First, it could be done through the addition of specialized “temporal processing cells”, which would be essentially represented as simple structures with small number of inputs, small number of outputs and small amount of memory. Such approach will be probably appropriate for the lower levels of the hierarchy. On higher levels of the hierarchy, the relations between spatio­temporal parts will be explicitly modeled in the similar way than the spatial relations between the parts in shape­ based object categorization example [Fidler2010a]. The introduction of temporal dimension into the model will require identification of the associated temporal parameters.

Finally, different levels of hierarchy will provide invariance to different variations of input data. This functionality will have to be adjusted properly, to achieve the degree of invariance that is desired in such a model.

To test the adequacy of our motion descriptors, influence of temporal component, invariance to different aspects of variation in the data, and to determine the rare­event regions in the parameter space, we will devise an activity recognition experiment. In this experiment we will use our motion descriptors in a standard classification problem on a publicly available activity recognition data­set [Schuldt2004] which will not only allow for analysis of our model, but will also provide comparison with the abundance of the related approaches for activity recognition that have been tested on the same dataset.

This website uses cookies to manage authentication, navigation, and other functions. By using our website, you agree that we can place these types of cookies on your device.

View e-Privacy Directive Documents

You have declined cookies. This decision can be reversed.

You have allowed cookies to be placed on your computer. This decision can be reversed.