Achieving robust and real-time 3D perception is fundamental for autonomous vehicles. While most existing 3D perception methods prioritize detection accuracy, they often overlook critical aspects such as computational efficiency, onboard chip deployment friendliness, resilience to sensor mounting deviations, and adaptability to various vehicle types. To address these challenges, we present NVAutoNet: a specialized Bird’s-Eye-View (BEV) perception network tailored explicitly for automated vehicles. NVAutoNet takes synchronized camera images as input and predicts 3D signals like obstacles, freespaces, and parking spaces. The core of NVAutoNet’s architecture (image and BEV backbones) relies on efficient convolutional networks, optimized for high performance using TensorRT. Our image-to-BEV transformation employs simple linear layers and BEV lookup tables, ensuring rapid inference speed. Trained on an extensive proprietary dataset, NVAutoNet consistently achieves elevated perception accuracy, operating remarkably at 53 frames per second on the NVIDIA DRIVE Orin SoC. Notably, NVAutoNet demonstrates resilience to sensor mounting deviations arising from diverse car models. Moreover, NVAutoNet excels in adapting to varied vehicle types, facilitated by inexpensive model fine-tuning procedures that expedite compatibility adjustments.

Synthetic data generation to improve classification performance (data augmentation) is a well-studied problem. Recently, generative adversarial networks (GAN) have shown superior image data augmentation performance, but their suitability in gesture synthesis has received inadequate attention. Further, GANs prohibitively require simultaneous generator and discriminator network training. We tackle both issues in this work. We first discuss a novel, device-agnostic GAN model for gesture synthesis called DeepGAN. Thereafter, we formulate DeepNAG by introducing a new differentiable loss function based on dynamic time warping and the average Hausdorff distance, which allows us to train DeepGAN’s generator without requiring a discriminator. Through evaluations, we compare the utility of DeepGAN and DeepNAG against two alternative techniques for training five recognizers using data augmentation over six datasets. We further investigate the perceived quality of synthesized samples via an Amazon Mechanical Turk user study based on the HYPE∞ benchmark. We find that DeepNAG outperforms DeepGAN in accuracy, training time (up to 17 × faster), and realism, thereby opening the door to a new line of research in generator network design and training for gesture synthesis. Our source code is available at https://www.deepnag.com.

We present Machete, a straightforward segmenter one can use to isolate custom gestures in continuous input. Machete uses traditional continuous dynamic programming with a novel dissimilarity measure to align incoming data with gesture class templates in real time. Advantages of Machete over alternative techniques is that our segmenter is computationally efficient, accurate, device-agnostic, and works with a single training sample. We demonstrate Machete’s effectiveness through an extensive evaluation using four new high-activity datasets that combine puppeteering, direct manipulation, and gestures. We find that Machete outperforms three alternative techniques in segmentation accuracy and latency, making Machete the most performant segmenter. We further show that when combined with a custom gesture recognizer, Machete is the only option that achieves both high recognition accuracy and low latency in a video game application.

Those who design gesture recognizers and user interfaces often use data collection applications that enable users to comfortably produce gesture training samples. In contrast, games present unique contexts that impact cognitive load and have the potential to elicit rapid gesticulations as players react to dynamic conditions, which can result in high gesture form variability. However, the extent to which these gestures differ is presently unknown. To this end, we developed two games with unique mechanics, Follow the Leader (FTL) and Sleepy Town, as well as a standard data collection application. We collected gesture samples from 18 participants across all conditions for gestures of varying complexity, and through an analysis using relative, global, and distribution coverage measures, we confirm significant differences between conditions. We discuss the implications of our findings, and show that our FTL design is closer to being an ecologically valid data collection protocol with low implementation complexity.

We introduce an interactive system for extracting the geometries of generalized cylinders and cuboids from single-or multiple-view point clouds. Our proposed method is intuitive and only requires the object's silhouettes to be traced by the user. Leveraging the user's perceptual understanding of what an object looks like, our proposed method is capable of extracting accurate models, even in the presence of occlusion, clutter or incomplete point cloud data, while preserving the original object's details and scale. We demonstrate the merits of our proposed method through a set of experiments on a public RGB-D dataset. We extracted 16 objects from the dataset using at most two views of each object. Our extracted models represent a high degree of visual similarity to the original objects. Further, we achieved a mean normalized Hausdorff distance of 5.66% when comparing our extracted models with the dataset's ground truths.

Despite decades of research, there is yet no general rapid prototyping recognizer for dynamic gestures that can be trained with few samples, work with continuous data, and achieve high accuracy that is also modality-agnostic. To begin to solve this problem, we describe a small suite of accessible techniques that we collectively refer to as the Jackknife gesture recognizer. Our dynamic time warping based approach for both segmented and continuous data is designed to be a robust, go-to method for gesture recognition across a variety of modalities using only limited training samples. We evaluate pen and touch, Wii Remote, Kinect, Leap Motion, and sound-sensed gesture datasets as well as conduct tests with continuous data. Across all scenarios we show that our approach is able to achieve high accuracy, suggesting that Jackknife is a capable recognizer and good first choice for many endeavors.

We introduce Code Park, a novel tool for visualizing codebases in a 3D game-like environment. Code Park aims to improve a programmer's understanding of an existing codebase in a manner that is both engaging and intuitive, appealing to novice users such as students. It achieves these goals by laying out the codebase in a 3D park-like environment. Each class in the codebase is represented as a 3D room-like structure. Constituent parts of the class (variable, member functions, etc.) are laid out on the walls, resembling a syntax-aware "wallpaper". The users can interact with the codebase using an overview, and a first-person viewer mode. We conducted two user studies to evaluate Code Park's usability and suitability for organizing an existing project. Our results indicate that Code Park is easy to get familiar with and significantly helps in code understanding compared to a traditional IDE. Further, the users unanimously believed that Code Park was a fun tool to work with.

Training gesture recognizers with synthetic data generated from real gestures is a well known and powerful technique that can significantly improve recognition accuracy. In this paper we introduce a novel technique called gesture path stochastic resampling (GPSR) that is computationally efficient, has minimal coding overhead, and yet despite its simplicity is able to achieve higher accuracy than competitive, state-of-the-art approaches. GPSR generates synthetic samples by lengthening and shortening gesture subpaths within a given sample to produce realistic variations of the input via a process of nonuniform resampling. As such, GPSR is an appropriate rapid prototyping technique where ease of use, understandability, and efficiency are key. Further, through an extensive evaluation, we show that accuracy significantly improves when gesture recognizers are trained with GPSR synthetic samples. In some cases, mean recognition errors are reduced by more than 70%, and in most cases, GPSR outperforms two other evaluated state-of-the-art methods.

Visual pattern recognition and classification is a challenging computer vision problem. Genetic programming has been applied towards automatic visual pattern recognition. One of the main factors in evolving effective classifiers is the suitability of the GP language for defining expressions for feature extraction and classification. This research presents a comparative study of a variety of GP languages suitable for classification. Four different languages are examined, which use different selections of image processing operators. One of the languages does block classification, which means that an image is classified as a whole by examining many blocks of pixels within it. The other languages are pixel classifiers, which determine classification for a single pixel. Pixel classifiers are more common in the GP-vision literature. We tested the languages on different instances of Brodatz textures, as well as aerial and camera images. Our results show that the most effective languages are pixel-based ones with spatial operators. However, as is to be expected, the nature of the image will determine the effectiveness of the language used.