Accurate tracking of targets is vital for safe and reliable operations, particularly in complex and dynamic environments such as urban areas. Traditional tracking methods, including Kalman and particle filters, often perform poorly in real world scenarios, due to inaccurate models and sparse or noisy measurements. Gaussian process (GP) based methods offer a flexible and data driven alternative with uncertainty quantification that does not depend on predefined dynamical equations. However, state of the art GP tracking approaches require expensive hyperparameter optimization, which limits their practicality for real time applications. In this work, we introduce a novel GP mixture based computationally efficient tracking method, which is capable of modeling complex system behavior and adapt to changing dynamics. Our proposed solution, named Multiple Model Recursive Gaussian Process (MM-RGP), adapts continuously to changing dynamics, is capable of modeling complex behavior, and is robust against sparse observation. In addition, the proposed method avoids hyperparameter optimization and adapts to incoming data. We demonstrate the effectiveness of our solution using the example of unmanned aerial vehicle (UAV) tracking, with both simulated and real datasets, and propose directions for extending our work.

Large annotated datasets are crucial for the success of deep learning, but labeling data can be prohibitively expensive in domains such as medical imaging. This work tackles the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">subset selection problem:</i> selecting a small set of the most informative examples from a large unlabeled pool for annotation. We propose a simple and effective method that combines feature norms, randomization, and orthogonality (via the Gram–Schmidt process) to select diverse and informative samples. Feature norms serve as a proxy for informativeness, while randomization and orthogonalization reduce redundancy and encourage coverage of the feature space. Extensive experiments on image and text benchmarks, including CIFAR-10/100, Tiny ImageNet, ImageNet, OrganAMNIST, and Yelp, show that our method consistently improves subset selection performance, both as a standalone approach and when integrated with existing techniques.

Binaural processing is becoming an important feature of high-end commercial headsets and hearing aids. Speech enhancement with binaural output requires adequate treatment of spatial cues in addition to desirable noise reduction and simultaneous speech preservation. Binaural speech enhancement was traditionally approached with model-based statistical signal processing, where the principle of common-gain filtering with identical treatment of left- and right-ear signals has been designed to achieve enhancement constrained by strict binaural cue preservation. However, model-based approaches may also be instructive for the design of modern deep learning architectures. In this article, the common-gain paradigm is therefore embedded into an artificial neural network approach. In order to maintain the desired common-gain property end-to-end, we derive the requirements for compressed feature formation and data normalization. Binaural experiments with moderate-sized artificial neural networks demonstrate the superiority of the proposed common-gain autoencoder network over model-based processing and related unconstrained network architectures for anechoic and reverberant noisy speech in terms of segmental SNR, binaural perception-based metrics MBSTOI, better-ear HASQI, and a listening experiment.

We present LEMON, a mesh editing pipeline that integrates neural deferred shading with localized mesh optimization to enable fast and precise editing of polygonal meshes guided by text prompts. Existing solutions for this problem tend to focus on a single task, either geometry or novel view synthesis, which often leads to disjointed results between the mesh and view. Our approach starts by identifying the most important vertices in the mesh for editing, using a segmentation model to focus on these key regions. Given multi-view images of an object, we optimize a neural shader and a polygonal mesh while extracting the normal map and the rendered image from each view. Using these outputs as conditioning data, we edit the input images with a text-to-image diffusion model and iteratively update our dataset while deforming the mesh. This process results in a polygonal mesh that is edited according to the given text instruction, preserving the geometric characteristics of the initial mesh while focusing on the most significant areas. We evaluate our pipeline on the DTU dataset, demonstrating that it generates finely-edited meshes more rapidly than the current state-of-the-art methods. We include our code and additional results in the supplementary material.

In [1, p. 1124], a footnote is needed on (13) as shown below: \begin{equation*}\qquad\qquad\quad{{\alpha }^\# } < \left( {1 - {{c}_1}} \right)\alpha + \left( {1 - \left( {1 - {{c}_1}} \right)\alpha } \right)\alpha\qquad\qquad\quad \hbox{(13)$^{1}$} \end{equation*}

We propose a novel snapshot hyperspectral imaging method that incorporates co-designed optics, a color filter array (CFA), and an unrolled post-processing network through end-to-end learning. The camera optics consist of a fixed refractive lens and a diffractive optical element (DOE). The learned DOE and CFA efficiently encode the hyperspectral data cube on the sensor via phase and amplitude modulation at the camera aperture and sensor planes, respectively. Subsequently, the unrolled network reconstructs the hyperspectral images from the sensor signal with high accuracy. We conduct extensive simulations to analyze and validate the performance of the proposed method for several CFA models and in non-ideal imaging conditions. We demonstrate that the Gaussian model is effective for parameterizing the spectral transmission functions of CFA pixels, providing high reconstruction accuracy and being relatively easy to implement. Furthermore, we show that learned CFA patterns are effective when optimally coupled with co-designed diffractive-refractive optics. We evaluate the robustness of our method against sensor noise and potential inaccuracies in the fabrication of the DOE and CFA. Our results show that our method achieves superior reconstruction quality compared to state-of-the-art methods, excelling in both spatial and spectral detail recovery and maintaining robustness against realistic noise levels. The code is available at the GitHub repository: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/ayoungkim-tech/Unrolled_C-DOE_HSI.git</uri>.

Detection and estimation performance depends on signal-to-interference-plus-noise ratio (SINR) at the output of an array. The Capon beamformer (BF) designed with ensemble statistics achieves the optimum SINR in stationary environments. Adaptive BFs compute their weights using the sample covariance matrix (SCM) obtained from snapshots, i.e., training samples. SINR loss, the ratio of adaptive to optimal SINR, quantifies the number of snapshots required to achieve a desired average level of performance. For adaptive Capon BFs that invert the full SCM, Reed et al. derived the SINR loss distribution and Miller quantified how the desired signal's presence in the snapshots degrades that loss. Abraham and Owsley designed dominant mode rejection (DMR) for cases where the number of snapshots is less than or approximately equal to the number of sensors. DMR's success in snapshot-starved passive sonar scenarios led to its application in other areas such as hyperspectral sensing and medical imaging. DMR forms a modified SCM as a weighted combination of the identity matrix and the dominant eigensubspace containing the loud interferers, thereby eliminating the inverse of the poorly estimated noise subspace. This work leverages recent random matrix theory (RMT) results to develop DMR performance predictions under the assumption that the desired signal is contained in the training data. Using white noise gain and interference suppression predictions, the paper derives a lower bound on DMR's average SINR loss and confirms its accuracy using Monte Carlo simulations. Moreover, this paper creates a new eigensubspace leakage estimator applicable to broader RMT applications.

Deep neural networks (DNNs) were shown to facilitate the operation of uplink multiple-input multiple-output (MIMO) receivers, with emerging architectures augmenting modules of classic receiver processing. Current designs employ static DNNs, whose architecture is fixed and weights are pre-trained. This poses a notable challenge, as the resulting MIMO receiver is suitable for a given configuration, i.e., channel distribution and number of users, while in practice these parameters change frequently with network variations and users leaving and joining the network. In this work, we tackle this core challenge of DNN-aided MIMO receivers. We build upon the concept of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">hypernetworks</i> , augmenting the receiver with a pre-trained deep model whose purpose is to update the weights of the DNN-aided receiver in response to instantaneous channel variations. We design our hypernetwork to augment <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">modular</i> deep receivers, leveraging their modularity to have the hypernetwork adapt not only the weights, but also the architecture. Our modular hypernetwork leads to a DNN-aided receiver whose architecture and resulting complexity adapt to the number of users, as well as to channel variations, without re-training. Our numerical studies demonstrate superior error-rate performance of modular hypernetworks in time-varying channels compared to static pre-trained receivers, while providing rapid adaptivity and scalability to network variations.

A method for performing spatial upsampling of Head-Related Impulse Responses (HRIRs) from sparse measurements is introduced. Based on a supervised elevation-wise encoder-decoder network design, we present two variants: one that performs progressive reconstructions with feed-forward connections from higher to lower elevations, and another that excludes these connections. The variants were evaluated in terms of the errors in interaural time and level differences, as well as the spectral distortion in the ipsilateral and contralateral ears. The additional complexity introduced by the variant with feed-forward connections does not always translate into accuracy gains, making the simpler variant preferable for efficiency. Performance generally improved as the number of available measurements increased. However, accuracy was also found to strongly depend on the spatial distribution of those measurements. Compared to an average non-personalized HRIRs, interaural time differences remain similar, while the proposed method achieves higher spectral and level accuracy, highlighting its practical use for HRIR upsampling.

We introduce Label-Combination Prototypical Networks (LC-Protonets) to address the problem of multi-label few-shot classification, where a model must generalize to new classes based on only a few available examples. Extending Prototypical Networks, LC-Protonets generate one prototype per label combination, derived from the power set of labels present in the limited training items, rather than one prototype per label. Our method is applied to automatic audio tagging across diverse music datasets, covering various cultures and including both modern and traditional music, and is evaluated against existing approaches in the literature. The results demonstrate a significant performance improvement in almost all domains and training setups when using LC-Protonets for multi-label classification. In addition to training a few-shot learning model from scratch, we explore the use of a pre-trained model, obtained via supervised learning, to embed items in the feature space. Fine-tuning improves the generalization ability of all methods, yet LC-Protonets achieve high-level performance even without fine-tuning, in contrast to the comparative approaches. We finally analyze the scalability of the proposed method, providing detailed quantitative metrics from our experiments. The implementation and experimental setup are made publicly available, offering a benchmark for future research.