Yuan Zeng

Neural Radiance Fields (NeRFs) have gained popularity by demonstrating impressive capabilities in synthesizing realistic novel views from multiple view images. It approximates the continuous integration of rays as a finite Riemann sum of the estimated colors and densities of the sampled points. Although this allows for efficient rendering, approximating divergent integrals under varying directions and interval lengths with piecewise constant features does not account for high-order variations within integration intervals, leading to ambiguous representations and limited reconstruction quality. In this paper, we propose to model the distribution within each interval using explicit Taylorlike explicit polynomials parameterized by spatial gradients. These polynomials encode the length and direction information of integrals, helping to disambiguate interval distributions and mitigate integral approximation errors in volume rendering. Our framework includes a learnable gradient estimator, regularized via an interval boundary loss to ensure numerical consistency and smoothness, alongside an adaptive interval scaling module that enhances stability and handles varying interval sizes. Our proposed method allows an easy integration with existing NeRFbased rendering frameworks. Experimental results on both synthetic and real-world scenes demonstrate that our method significantly boosts the rendering quality of various NeRF models, achieving state-of-the-art performance.

In wireless communication systems, the modulated radio signals are commonly susceptible to sensor-level changes and defects during transmission, resulting in data distribution shifts in the received signals. Deep learning-based automatic modulation recognition (AMR) has made significant strides but still struggles with performance on out-of-distribution (OOD) samples due to domain shifts in practical communication systems. Test-time adaptation (TTA) methods adapt models using test samples at inference time, emerging as a promising solution to this challenge. In this letter, we propose a novel self-supervised TTA strategy to adapt mask autoencoders to better recognize modulation modes in OOD scenarios. The key idea is to optimize the main modulation recognition task and the modulated signal spectrogram reconstruction task during training toward effective feature extraction, and design an entropy control mechanism to adapt the model toward better modulation recognition of the test signals. With extensive experiments, we show that the proposed method effectively improves the performance of TTA for AMR in various distribution shift scenarios.

3D scene stylization refers to generating stylized images of the scene at arbitrary novel view angles following a given set of style images while ensuring consistency when rendered from different views. Recently, several 3D style transfer methods leveraging the scene reconstruction capabilities of pre-trained neural radiance fields (NeRF) have been proposed. To successfully stylize a scene this way, one must first reconstruct a photo-realistic radiance field from collected images of the scene. However, when only sparse input views are available, pre-trained few-shot NeRFs often suffer from high-frequency artifacts, which are generated as a by-product of high-frequency details for improving reconstruction quality. Is it possible to generate more faithful stylized scenes from sparse inputs by directly optimizing encoding-based scene representation with target style? In this paper, we consider the stylization of sparse-view scenes in terms of disentangling content semantics and style textures. We propose a coarse-to-fine sparse-view scene stylization framework, where a novel hierarchical encoding-based neural representation is designed to generate high-quality stylized scenes directly from implicit scene representations. We also propose a new optimization strategy with content strength annealing to achieve realistic stylization and better content preservation. Extensive experiments demonstrate that our method can achieve high-quality stylization of sparse-view scenes and outperforms fine-tuning-based baselines in terms of stylization quality and efficiency.

Neural radiance fields (NeRFs) have achieved impressive view synthesis results by learning an implicit volumetric representation from multi-view images. To project the implicit representation into an image, NeRF employs volume rendering that approximates the continuous integrals of rays as an accumulation of the colors and densities of the sampled points. Although this approximation enables efficient rendering, it ignores the direction information in point intervals, resulting in ambiguous features and limited reconstruction quality. In this paper, we propose a learning method that utilizes learnable view-dependent features to improve scene representation and reconstruction. We model the volume rendering integral with a piecewise constant volume density and spherical harmonic-guided view-dependent features, facilitating ambiguity elimination while preserving the rendering efficiency. In addition, we introduce a regularization term that restricts the anisotropic representation effect to be local, with negligible effect on geometry representations, and that encourages recovering the correct geometry. Our method is flexible and can be plugged into NeRF-based frameworks. Extensive experiments show that the proposed representation can boost the rendering quality of various NeRFs and achieve state-of-the-art rendering performance on both synthetic and real-world scenes.

Radio frequency (RF) fingerprinting is a promising technique for device authentication due to its advantages of difficult-to-tamper and inimitableness. Rather than employing hand-crafted features, deep learning-based methods that learn low-level signal representation from raw signals have been explored recently. This letter proposes a novel RF fingerprinting scheme using learnable short-time Fourier transform (STFT) and convolutional neural network (CNN). Instead of representing radio signals by spectrograms from a fixed STFT and processing signal representation and classification separately, this approach integrates a parameterized STFT-based signal representation module and a CNN classifier into a single framework. The signal representation module is jointly trained with the CNN classifier using a single identification loss, transforming the input radio signals into spectrograms desired by the CNN classifier. Experimental results show that the RF fingerprint identification accuracy of the proposed scheme with learnable signal representation is significantly improved compared to baseline schemes with traditional hand-crafted signal representations.

Recent advances in Neural radiance fields (NeRF) have enabled high-fidelity scene reconstruction for novel view synthesis. However, NeRF requires hundreds of network evaluations per pixel to approximate a volume rendering integral, making it slow to train. Caching NeRFs into explicit data structures can effectively enhance rendering speed but at the cost of higher memory usage. To address these issues, we present Hyb-NeRF, a novel neural radiance field with a multi-resolution hybrid encoding that achieves efficient neural modeling and fast rendering, which also allows for high-quality novel view synthesis. The key idea of Hyb-NeRF is to represent the scene using different encoding strategies from coarse-to-fine resolution levels. Hyb-NeRF exploits memory-efficiency learnable positional features at coarse resolutions and the fast optimization speed and local details of hash-based feature grids at fine resolutions. In addition, to further boost performance, we embed cone tracing-based features in our learnable positional encoding that eliminates encoding ambiguity and reduces aliasing artifacts. Extensive experiments on both synthetic and real-world datasets show that Hyb-NeRF achieves faster rendering speed with better rending quality and even a lower memory footprint in comparison to previous state-of-the-art methods.

Recently, channel state information (CSI) extracted from Wi-Fi signals has enabled a variety of Wi-Fi sensing applications beyond communications. However, the extraction of CSI relies on specific Wi-Fi devices, which severely limits the CSI-based sensing applications. In this paper, we exploit the beamforming matrix, a compressed version of CSI that is fed back from the beamformee to the beamformer without encryption, for Wi-Fi sensing. In order to recover more channel information from the beamforming matrix for sensing, a QR decomposition (QRD)-based beamforming feedback scheme is proposed in this paper. We analyze the effectiveness of the proposed QRD-based scheme in terms of sensing and communication performance. Simulation results show that the Doppler frequency shift, time of flight, angle of departure, as well as the amplitude information can be recovered using the proposed scheme and at the same time, the spectral efficiency of the proposed QRD-based scheme is comparable to that of the existing SVD-based scheme. We implement the proposed scheme on IEEE 802.11ac/ax based Wi-Fi devices. The experimental results on human respiration and finger tracking demonstrate that the proposed scheme works well for practical Wi-Fi sensing.

Radio frequency (RF) fingerprinting is a promising device authentication technique for securing the Internet of Things. It exploits the intrinsic and unique hardware impairments of the transmitters for device identification. Recently, due to the superior performance of deep learning (DL)-based classification models on real-world datasets, DL networks have been explored for RF fingerprinting. Most existing DL-based RF fingerprinting models use a single representation of radio signals as the input, while the multi-channel input model can leverage information from different representations of radio signals and improve the identification accuracy of RF fingerprints. In this work, we propose a multi-channel attentive feature fusion (McAFF) method for RF fingerprinting. It utilizes multi-channel neural features extracted from multiple representations of radio signals, including in-phase and quadrature (IQ) samples, carrier frequency offsets, fast Fourier transform coefficients and short-time Fourier transform coefficients. The features extracted from different channels are fused adaptively using a shared attention module, where the weights of neural features are learned during the model training. In addition, we design a signal identification module using a convolution-based ResNeXt block to map the fused features to device identities. To evaluate the identification performance of the proposed method, we construct a Wi-Fi dataset, named Wi-Fi Device Identification (WFDI), using commercial Wi-Fi end-devices as the transmitters and a Universal Software Radio Peripheral (USRP) platform as the receiver. Experimental results show that the proposed McAFF method significantly outperforms the single-channel-based as well as the existing DL-based RF fingerprinting methods in terms of identification accuracy and robustness.