We optimized and deployed the adaptive framework Virtuoso that can maintain real-time object detection even when experiencing high contention scenarios. The original Virtuoso framework uses an adaptive algorithm for the detection frame followed by a low-cost algorithm for the tracker frame which uses down-sampled images to reduce computation. One of our optimizations include detaching the single synchronous thread for detection and tracking into two parallel threads. This multi-threaded implementation allows for computationally high-cost detection algorithms to be used while still maintaining real-time output from the tracker thread. Another optimization we developed uses multiple down-sampled images to initialize each tracker based on the size of the input box; the multiple down-sampled images allow each tracker to choose the optimal image size for the box that it is tracking rather than a single down-sampled image being used for all trackers.
Supervised machine learning depends on training a model to mimic previous labeled results. The problem with a small dataset is that data augmentation is necessary to increase the generalization of the model to future images, but we have observed that future images won’t necessarily be in the same domain as the augmented images. To alleviate this problem we applied image segmentation multiple times on the same image by using the same data augmentation techniques on the image in question, and then we merged the results using a priority based on the class weights used when training the model. Merging the segmentation results from the augmented images increased the mean-intersection-over-union over the inference results that used a single image.
With the advent of neural networks, users at the tactical edge have started experimenting with AI enabled intelligent mission applications. Autonomy stacks have been proposed for the tactical environments for sensing, reasoning and computing the situational awareness to provide the human in the loop actionable intelligence in mission time. Tactical edge computing platforms must employ small-form-factor modules for compute, storage, and networking functions that conform to strict size, weight, and power constraints (SWaP). Many of the neural network models proposed for the tactical AI stack are computationally complex and may not be deployable without modifications. In this paper we discuss deep neural network optimization approaches for resource constrained tactical unmanned ground vehicles.
Normalizations used in model reduction can be chosen to emphasize anything from computation reduction to parameter reduction. Choosing a normalization that emphasizes a model with a small number of parameters is useful when deploying a model onto machines with a limited communication rate, while choosing a normalization that emphasizes a model with a small computational cost is useful when deploying a model onto a machine for real-time sensor analysis. As such, we explore the effect of various normalizations used to prune kernel parameters on models trained on the ImageNet database.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.