Token-based masked generative models are gaining popularity for their fast inference time with parallel decoding. While recent token-based approaches achieve competitive performance to diffusion-based models, their generation performance is still suboptimal as they sample multiple tokens simultaneously without considering the dependence among them. We empirically investigate this problem and propose a learnable sampling model, Text-Conditioned Token Selection (TCTS), to select optimal tokens via localized supervision with text information. TCTS improves not only the image quality but also the semantic alignment of the generated images with the given texts. To further improve the image quality, we introduce a cohesive sampling strategy, Frequency Adaptive Sampling (FAS), to each group of tokens divided according to the self-attention maps. We validate the efficacy of TCTS combined with FAS with various generative tasks, demonstrating that it significantly outperforms the baselines in image-text alignment and image quality. Our text-conditioned sampling framework further reduces the original inference time by more than 50 % without modifying the original generative model.

Automated understanding and recognition of human activities and behaviors in a smart space (e.g., smart house) is of paramount importance to many critical human-centered applications. Recognized activities are the input to the pervasive computer (the smart space) which intelligently interacts with the users to maintain the application's goal be it assistance, safety, child-development, entertainment or other goals. Research in this area is fascinating but severely lacks adequate validation which often relies on datasets that contain sensory data representing the activities. Availing adequate datasets that can be used in a large variety of spaces, for different user groups, and aiming at different goals is very challenging. This is due to the prohibitive cost and the human capital needed to instrument physical spaces and to recruit human subjects to perform the activities and generate data. Simulation of human activities in smart spaces has therefore emerged as an alternative approach to bridge this deficit. Traditional event-driven approaches have been proposed. However, the complexity of human activity simulation was proved to be challenging to these initial simulation efforts. In this paper, we present Persim 3D-an alternative context-driven approach to simulating human activities capable of supporting complex activity scenarios. We present the context-activity-action nexus and show how our approach combines modeling and visualization of actions with context and activity simulation. We present the Persim 3D architecture and algorithms, and describe a detailed validation study of our approach to verify the accuracy and realism of the simulation output (datasets and visualizations) and the scalability of the human effort in using Persim 3D to simulate complex scenarios. We show positive and promising results that validate our approach.

As demands for human activity recognition technology increase, simulation of human activities for providing datasets and testing purposes is becoming increasingly important. Traditional simulation, however, is based on an event-driven approach, which focuses on single sensor events and models within a single human activity. It requires detailed description and processing of every low-level event that enters into an activity scenario. For many realistic and complex human scenarios, the event-driven approach burdens the simulator users with complicated low-level specifications required to configure and run the simulation. It also increases computational complexity and impedes scalable simulation. Thus, we propose a novel, context-driven approach to simulating human activities in smart spaces. In the proposed approach, vectors of sensors rather than single sensor events drive the simulation quicker from one context to another. Abstracting the space state into contexts highly simplifies the tasks and efforts of the simulation user in setting up and configuring the simulation components for smart space and human activities. We present the context-driven simulation approach and show how it works. Then we present fundamental concepts and algorithms and provide a comparative performance study between the event- and context-driven simulation approaches.

Emerging smart space applications are increasingly relying on capabilities for recognizing human activities. Activity recognition research is however challenged and slowed by the lack of data necessary for testing and validation. Collecting data through live-in trials in real world deployments is often very expensive and complicated. Legitimate limitations on the use of human subjects also renders a much smaller dataset than desired to be collected. To address this challenge, we propose a scenario generation approach in which a small set of scenarios is used to generate new relevant and realistic scenarios, and hence increase the base of testing data needed for activity recognition validation. Unlike existing methods for generating scenarios, which usually focus on scenario structure and complexity, we propose a Bayesian-based approach that learns the stochastic characteristics of a small number of collected datasets to generate additional scenarios of similar characteristics. Our approach is prolific and can generate enormous datasets with high degree of realism at affordable cost. The proposed approach is validated using a Viterbi-based algorithm and a real dataset case study. The validation experiment confirms that the generated dataset has highly similar stochastic characteristics as that of the real dataset.

Activity recognition research relies heavily on test data to verify the modeling technique and the performance of the activity recognition algorithm. But data from real deployments are expensive and time consuming to obtain. And even if cost is not an issue, regulatory limitations on the use of human subjects prohibit the collection of extensive datasets that can test all scenarios, under all circumstances. A powerful and verifiable simulation tool is needed to accelerate research on human activity recognition. We present Persim, an event driven simulator of human activities in pervasive spaces. Persim is capable of capturing elements of space, sensors, behaviors (activities), and their inter-relationships. We focus on presenting the five main use cases for Persim addressing dataset synthesis, reuse and extension of existing datasets, sharing of data and simulation projects, as well as data validation.

The generation of useful sensory data from real-world deployments of Ambient Intelligent Environments (AIEs) is challenging because of the high cost, significant groundwork and lack of access to human subjects. This situation can be improved by providing efficient simulators that can produce realistic simulation of the data collection from AIEs. One of the main problems for developing AIE simulators lies in the ability to verify how close the simulated data are to the real world data. In this paper, we present a fuzzy based verification agent for Persim - an event driven simulator for human activities in AIEs. The employed fuzzy based verification agent builds a data model that mimics the operation of Persim which allows for the latter's objective and subjective verification. We have conducted the verification on real world data captured from an actual smart apartment deployment. The results show the effectiveness of the fuzzy based verification agent in analyzing and comparing the Persim simulated data with the real world collected data. We also demonstrate how the verification agent is able to pinpoint specific changes to the simulation model to increase the realism of the simulation.