Human Activity Recognition (HAR) is widely applied on wearable devices in our daily lives. However, acquiring high-quality wearable sensor data set with ground-truths is challenging due to the high cost in collecting data and necessity of domain experts. In order to achieve generalization from limited data, we study augmentation-based Self-Supervised Learning (SSL) for data from wearable devices. However, there is an issue in one of the most popular SSL approaches, contrastive learning: it is sensitive to the choice of data augmentations. To resolve this, we first propose to combine contrastive learning with generative learning, which is robust to augmentations. Second, we propose an automatic augmentation policy search method to discover the most promising augmentation policy. We empirically verify our approaches on three public HAR datasets. Experimental results show that our proposed SSL approach is robust to augmentations, and delivers higher accuracy than contrastive learning. Additionally, with the searched augmentation policy we are able to further improve the accuracy of HAR task.

As one of the successful Transformer-based models in computer vision tasks, SegFormer demonstrates superior performance in semantic segmentation. Nevertheless, the high computational cost greatly challenges the deployment of SegFormer on edge devices. In this paper, we seek to design a lightweight SegFormer for efficient semantic segmentation. Based on the observation that neurons in SegFormer layers exhibit large variances across different images, we propose a dynamic gated linear layer, which prunes the most uninformative set of neurons based on the input instance. To improve the dynamically pruned SegFormer, we also introduce two-stage knowledge distillation to transfer the knowledge within the original teacher to the pruned student network. Experimental results show that our method can significantly reduce the computation overhead of SegFormer without an apparent performance drop. For instance, we can achieve 36.9% mIoU with only 3.3G FLOPs on ADE20K, saving more than 60% computation with the drop of only 0.5% in mIoU.

The concept of Riemannian geometries is used to construct induced homogeneous point sets on manifolds that are based on well-distributed point sets in unit cubes of an appropriately chosen Euclidean space. These well-distributed point sets in unit cubes are based on standard low-discrepancy sequences. The approach is algorithmic, that is, the methods developed in this article have been implemented and tested. Applications in image processing, graph theory and measurement-based exploration are presented.

Automatic target detection and recognition (ATD/R) is of crucial interest to the defense community. We present a robust ATD/R system developed at the CVRC at UT-Austin for recognition in second generation forward looking infrared (FLIR) images. An experiment conducted on 1930 FLIR images shows that this ATR system can achieve recognition with a high degree of accuracy and a low false alarm rate. This demo first presents a brief overview of the whole methodology, then shows the detailed procedures and temporary outputs step by step, by running this ATR system on typical low-contrast FLIR images. Results and examples are presented at the end of the demonstration.

This paper presents a system that detects unexpected moving obstacles that appear in the path of a navigating robot, and estimates the relative motion of the object with respect to the robot. The system is designed for a robot navigating in a structured environment with a single wide-angle camera. The system uses polar mapping to simplify the segmentation of the moving object from the background. The polar mapping is performed with the focus of expansion as the center. A vision-based algorithm that uses the vanishing points of segments extracted from a scene in a few 3D orientations provides an accurate estimate of the robot orientation. In the transformed space qualitative estimate of moving obstacles is obtained by detecting the vertical motion of edges extracted in a few specified directions. Relative motion information about the obstacle is then obtained by computing the time to impact between the obstacles and robot from the radial component of the optical flow. The system was implemented and on an indoor mobile robot.

We present a methodology for object recognition by parts. The methodology is based on a hierarchical, modular structure for object recognition. Recognition is performed at different levels in the hierarchy, and the type of recognition performed differs from level to level. Each level is made up of modules, where each module is an expert on a particular part of an object, that is, each module is specifically trained to recognize one part of an object. We present a Bayesian system in which the expert modules represent the probability density functions of each part, modeled as a mixture of densities to incorporate different views (aspects) of each part. Results obtained for object recognition in second generation forward looking infrared (FLIR) images are also presented in this paper.

In this paper we present a system for the detection and recognition of targets in second generation forward looking infrared (FLIR) images. The system uses new algorithms for target detection and segmentation of the targets. Recognition is based on a methodology far target recognition by parts. A diffusion based approach for determining the parts of a target is also presented here. Experimental results on a large database of FLIR images validate the robustness of the system, and its applicability to FLIR imagery obtained from real scenes.

We give a methodology for the evaluation and comparison of object recognition systems. The methodology is based on indicators of two kinds: (1) statistical and (2) algorithmic. Statistical indicators measure the significance of the performance difference between different systems and rank the performances when the difference is significant. Of the various statistical indicators, we use the Kruskal-Wallis H test. Algorithmic indicators include the usual space and time complexity measures, and various performance curves in variables of error and test data sample size. We illustrate the methodology to evaluate a number of pattern classifiers.

Presents a system that detects unexpected moving obstacles that appear in the path of a navigating robot, and that estimates the relative motion of the object with respect to the robot. The system is designed for a robot navigating in a structured environment with a single wide-angle camera. The system uses polar mapping to simplify the segmentation of the moving object from the background. The polar mapping is performed with the focus of expansion (FOE) as the center, obtained from the vanishing point of the significant lines in the environment. A method for determining a qualitative estimate of the motion of the object, coupled with an efficient clustering algorithm, is used to segment the moving obstacle from the background. The system uses an area based feature matching method to compute the optical flow in the segmented regions of the image sequence, and time-to-impact is computed from the radial component of the optical flow in the image plane. A framework for a real time implementation of this system on a robot is also presented.>