Despite all recent progress, it is still challenging to edit and manipulate natural images with modern generative models. When using Generative Adversarial Network (GAN), one major hurdle is in the inversion process mapping a real image to its corresponding noise vector in the latent space, since it is necessary to be able to reconstruct an image to edit its contents. Likewise for Denoising Diffusion Implicit Models (DDIM), the linearization assumption in each inversion step makes the whole deterministic inversion process unreliable. Existing approaches that have tackled the problem of inversion stability often incur in significant trade-offs in computational efficiency. In this work we propose an Accelerated Iterative Diffusion Inversion method, dubbed AIDI, that significantly improves reconstruction accuracy with minimal additional overhead in space and time complexity. By using a novel blended guidance technique, we show that effective results can be obtained on a large range of image editing tasks without large classifier-free guidance in inversion. Furthermore, when compared with other diffusion inversion based works, our proposed process is shown to be more robust for fast image editing in the 10 and 20 diffusion steps’ regimes.

Neural radiance fields (NeRF) have garnered significant attention, with recent works such as Instant-NGP accelerating NeRF training and evaluation through a combination of hashgrid-based positional encoding and neural networks. However, effectively leveraging the spatial sparsity of 3D scenes remains a challenge. To cull away unnecessary regions of the feature grid, existing solutions rely on prior knowledge of object shape or periodically estimate object shape during training by repeated model evaluations, which are costly and wasteful. To address this issue, we propose HollowNeRF, a novel compression solution for hashgrid-based NeRF which automatically sparsifies the feature grid during the training phase. Instead of directly compressing dense features, HollowNeRF trains a coarse 3D saliency mask that guides efficient feature pruning, and employs an alternating direction method of multipliers (ADMM) pruner to sparsify the 3D saliency mask during training. By exploiting the sparsity in the 3D scene to redistribute hash collisions, HollowNeRF improves rendering quality while using a fraction of the parameters of comparable state-of-the-art solutions, leading to a better cost-accuracy trade-off. Our method delivers comparable rendering quality to Instant-NGP, while utilizing just 31% of the parameters. In addition, our solution can achieve a PSNR accuracy gain of up to 1dB using only 56% of the parameters.

Mainstream video analytics uses a pre-trained DNN model with an assumption that inference input and training data follow the same probability distribution. However, this assumption does not always hold in the wild: autonomous vehicles may capture video with varying brightness; unstable wireless bandwidth calls for adaptive bitrate streaming of video; and, inference servers may serve inputs from heterogeneous IoT devices/cameras. In such situations, the level of input distortion changes rapidly, thus reshaping the probability distribution of the input. We present GearNN, an adaptive inference architecture that accommodates DNN inputs with varying distortions. GearNN employs an optimization algorithm to identify a tiny set of "distortion-sensitive" DNN parameters, given a memory budget. Based on the distortion level of the input, GearNN then adapts only the distortion-sensitive parameters, while reusing the rest of constant parameters across all input qualities. In our evaluation of DNN inference with dynamic input distortions, GearNN improves the accuracy (mIoU) by an average of 18.12% over a DNN trained with the undistorted dataset and 4.84% over stability training from Google, with only 1.8% extra memory overhead.

High-speed rail (HSR) systems potentially provide a more efficient way of door-to-door transportation than airplane. However, they also pose unprecedented challenges in delivering seamless Internet service for on-board passengers. In this paper, we conduct a large-scale active-passive measurement study of TCP performance over LTE on HSR. Our measurement targets the HSR routes in China operating at above 300 km/h. We performed extensive data collection through both controlled setting and passive monitoring, obtaining 1732.9 GB data collected over 135719 km of trips. Leveraging such a unique dataset, we measure important performance metrics such as TCP goodput, latency, loss rate, as well as key characteristics of TCP flows, application breakdown, and users' behaviors. We further quantitatively study the impact of frequent cellular handover on HSR networking performance, and conduct in-depth examination of the performance of two widely deployed transport-layer protocols: TCP CUBIC and TCP BBR. Our findings reveal the performance of today's commercial HSR networks "in the wild'', as well as identify several performance inefficiencies, which motivate us to design a simple yet effective congestion control algorithm based on BBR to further boost the throughput by up to 36.5%. They together highlight the need to develop dedicated protocol mechanisms that are friendly to extreme mobility.

IoT and deep learning based computer vision together create an immense market opportunity, but running deep neural networks (DNNs) on resource-constrained IoT devices remains challenging. Offloading DNN inference to an edge server is a promising solution, but limited wireless bandwidth bottlenecks its end-to-end performance and scalability. While IoT devices can adopt source compression to cope with the limited bandwidth, existing compression algorithms (or codecs) are not designed for DNN (but for human eyes), and thus, suffer from either low compression rates or high DNN inference errors. This paper presents GRACE, a DNN-aware compression algorithm that facilitates the edge inference by significantly saving the network bandwidth consumption without disturbing the inference performance. Given a target DNN, GRACE (i) analyzes DNN's perception model w.r.t both spatial frequencies and colors and (ii) generates an optimized compression strategy for the model -- one-time offline process. Next, GRACE deploys thus-generated compression strategy at IoT devices (or source) to perform online source compression within the existing codec framework, adding no extra overhead. We prototype GRACE on JPEG (the most popular image codec framework), and our evaluation results show that GRACE indeed achieves the superior compression performance over existing strategies for key DNN applications. For semantic segmentation tasks, GRACE reduces a source size by 23% compared to JPEG with similar interference accuracy (0.38% lower than GRACE). Further, GRACE even achieves 7.5% higher inference accuracy than JPEG with a commonly used quality level of 75 does. For classification tasks, GRACE reduces the bandwidth consumption by 90% over JPEG with the same inference accuracy.

Despite the pervasive use of real-time video telephony services, the users' quality of experience (QoE) remains unsatisfactory, especially over the mobile Internet. Previous work studied the problem via controlled experiments, while a systematic and in-depth investigation in the wild is still missing. To bridge the gap, we conduct a large-scale measurement campaign on \appname, an operational mobile video telephony service. Our measurement logs fine-grained performance metrics over 1 million video call sessions. Our analysis shows that the application-layer video codec and transport-layer protocols remain highly uncoordinated, which represents one major reason for the low QoE. We thus propose \name, a machine learning based framework to resolve the issue. Instead of blindly following the transport layer's estimation of network capacity, \name reviews historical logs of both layers, and extracts high-level features of codec/network dynamics, based on which it determines the highest bitrates for forthcoming video frames without incurring congestion. To attain the ability, we train \name with the aforementioned massive data traces using a custom-designed imitation learning algorithm, which enables \name to learn from past experience. We have implemented and incorporated \name into \appname. Our experiments show that \name outperforms state-of-the-art solutions, improving video quality while reducing stalling time by multi-folds under various practical scenarios.

Despite the pervasive use of real-time video telephony services, their quality of experience (QoE) remains unsatisfactory, especially over the mobile Internet. We conduct a large-scale measurement campaign on \appname, an operational mobile video telephony service. Our analysis shows that the application-layer video codec and transport-layer protocols remain highly uncoordinated, which represents one major reason for the low QoE. We thus propose \name, a machine learning based framework to resolve the issue. We train \name with the massive data traces from the measurement campaign using a custom-designed imitation learning algorithm, which enables \name to learn from past experience following an expert's iterative demonstration/supervision. We have implemented and incorporated \name into the \appname. Our experiments show that \name outperforms state-of-the-art solutions, improving video quality while reducing stalling time by multi-folds under various practical scenarios.

In recent years, augmented reality (AR) has drawn significant attention in the automotive industry and shown great potential for a variety of driver-assistance applications. Tracking the driver's head is vital to seamlessly merge the AR content in the driver's view but still remains an open problem. Specifically, most existing in-vehicle AR solutions rely on cameras for head tracking, which suffer from low sampling rate, weak performance at night, and even high computation costs. Wearing a dedicated headset, on the other hand, is intrusive and inconvenient for daily driving.

Panoramic or 360° video streaming has been supported by a wide range of content providers and mobile devices. Yet existing work primarily focused on streaming on-demand 360° videos stored on servers. In this paper, we examine a more challenging problem: Can we stream real-time interactive 360° videos across existing LTE cellular networks, so as to trigger new applications such as ubiquitous 360° video chat and panoramic outdoor experience sharing? To explore the feasibility and challenges underlying this vision, we design POI360, a portable interactive 360° video telephony system that jointly investigates both panoramic video compression and responsive video stream rate control. For the challenge that the legacy spatial compression algorithms for 360° video suffer from severe quality fluctuations as the user changes her region-of-interest (ROI), we design an adaptive compression scheme, which dynamically adjusts the compression strategy to stabilize the video quality within ROI under various user input and network condition. In addition, to meet the responsiveness requirement of panoramic video telephony, we leverage the diagnostic statistics on commodity phones to promptly detect cellular link congestion, hence significantly boosting the rate control responsiveness. Extensive field tests for our real-time POI360 prototype validate its effectiveness in enabling panoramic video telephony over the highly dynamic cellular networks.

Despite the 4G LTE's 10X capacity improvement over 3G, mobile Web loading latency remains a major issue that hampers user experience. The root cause lies in the inefficient transport-layer that underutilizes LTE capacity, due to high channel dynamics, wireless link losses, and insufficient application traffic to propel the bandwidth probing. In this paper, we propose Cellular Link-Aware Web loading (CLAW), which boosts mobile Web loading using a physical-layer informed transport protocol. CLAW harnesses the limited PHY-layer statistics available on LTE phones to quantitatively model the LTE channel resource utilization, which is then translated into a transport window that best fits the bandwidth. Consequently, CLAW can estimate and fully utilize the available bandwidth almost within one RTT. In addition, CLAW can precisely differentiate LTE wireless loss from congestion loss, and identify the rare cases when the wireline backhaul becomes the bottleneck. We have prototyped CLAW on commodity LTE phones. Across a wide range of experimental settings, CLAW consistently reduces Web loading latency by more than 30%, compared to classical TCP variants and state-of-the-art congestion controls for cellular networks.

Large-scale multiuser MIMO (MU-MIMO) systems have the potential for multi-fold scaling of network capacity. The research community has recognized this theoretical potential and developed architectures [1,2] with large numbers of RF chains. Unfortunately, building the hardware with a large number of RF chains is challenging in practice. CSI data transport and computational overhead of MU-MIMO beamforming can also become prohibitive under large network scale. Furthermore, it is difficult to physically append extra RF chains on existing communication equipments to support such large-scale MU-MIMO architectures.

Adaptive HTTP video streaming over LTE has been gaining popularity due to LTE's high capacity. Quality of adaptive streaming depends highly on the accuracy of client's estimation of end-to-end network bandwidth, which is challenging due to LTE link dynamics. In this paper, we present piStream, that allows a client to efficiently monitor the LTE basestation's PHY-layer resource allocation, and then map such information to an estimation of available bandwidth. Given the PHY-informed bandwidth estimation, piStream uses a probabilistic algorithm to balance video quality and the risk of stalling, taking into account the burstiness of LTE downlink traffic loads. We conduct a real-time implementation of piStream on a software-radio tethered to an LTE smartphone. Comparison with state-of-the-art adaptive streaming protocols demonstrates that piStream can effectively utilize the LTE bandwidth, achieving high video quality with minimal stalling rate.

Equipped with the multi-user MIMO (MU-MIMO) technology, a WiFi access point (AP) with M antennas can achieve M degrees-of-freedom (DoF) in theory. Existing MU-MIMO protocols strive to maximize DoF usage by serving M users simultaneously. In this paper, through a MU-MIMO testbed measurement, we found that the correlation between users' channels can severely compromise network throughput under a DoF-maximizing strategy. To combat this problem, we propose Kardia that judiciously distributes the DoF to least-correlated users within an AP's cell, and coordinates neighboring APs such that each can serve the best set of users while nulling mutual interference. The key challenge in Kardia lies in efficiently determining which set of users to serve without knowing the channel state from all of them. We propose a lightweight mechanism that enables APs to collaboratively infer the correlation between users through precoded probing packets. Our analysis shows that Kardia's DoF distribution framework has a provable performance guarantee. Our testbed implementation further reveals that Kardia can significantly boost the performance of multi-cell MU-MIMO networks, in contrast to legacy protocols like 802.11ac.

Full-duplex radios are often envisioned to double wireless link capacity. Substantial work has focused on redesigning the radio hardware to achieve this theoretical gain. From a network-protocol perspective, however, it remains an open problem how to exploit full-duplex radio, and how much gain it can achieve in practical multi-cell wireless LANs. In this paper, we propose FuMAC, a channel access protocol tailored for full-duplex radios to optimally exploit their unique capabilities. FuMAC addresses a unique trade off between PHY-layer full duplex transmission and MAC-level spatial reuse, through a semi synchronous channel access principle. Its design is enabled by a novel self-interference cancellation mechanism called Active Antenna Cancellation. We verify FuMAC using software-radio implementation combined with large scale simulation. The results demonstrate that conventional MAC protocols severely underutilize full-duplex's potential. In contrast, FuMAC can achieve more-than-doubled throughput gain over half-duplex wireless LANs and significantly outperform alternative full-duplex MAC designs, while maintaining a much higher level of fairness.

Full-duplex has emerged as a new communication paradigm and is anticipated to double wireless capacity. Existing studies of full-duplex mainly focused on its PHY layer design, which enables bidirectional transmission between a single pair of nodes. In this paper, we establish an analytical framework to quantify the network-level capacity gain of full-duplex over half-duplex. Our analysis reveals that inter-link interference and spatial reuse substantially reduces full-duplex gain, rendering it well below 2 in common cases. More remarkably, the asymptotic gain approaches 1 when interference range approaches transmission range. Through a comparison between optimal half- and full-duplex MAC algorithms, we find that full-duplex's gain is further reduced when it is applied to CSMA based wireless networks. Our analysis provides important guidelines for designing full-duplex networks. In particular, network-level mechanisms such as spatial reuse and asynchronous contention must be carefully addressed in full-duplex based protocols, in order to translate full-duplex's PHY layer capacity gain into network throughput improvement.

In a multi-user MIMO (MU-MIMO) network, an AP with M antennas can only serve up to M users out of a large user population. The M users' rates are inter-coupled and depend on their channel orthogonality. Substantial theoretical studies focused on selecting users to maximize capacity, but they require feedback of channel state information (CSI) from all users. The resulting overhead can easily overwhelm useful data in large scale networks. In this paper, we propose a scalable user selection mechanism called orthogonality probing based user selection (OPUS). OPUS only requires up to M rounds of CSI feedback. In each round, it employs a novel probing mechanism that enables a user to evaluate its orthogonality with existing users, and a distributed contention mechanism that singles out the best user to feedback its CSI. Software-radio based implementation and experimentation shows that OPUS significantly outperforms traditional user selection schemes in both throughput and fairness.

MIMO beamforming technology can scale wireless data rate proportionally with the number of antennas. However, the overhead induced by receivers' CSI (channel state information) feedback scales at a higher rate. In this paper, we address this fundamental tradeoff with Adaptive Feedback Compression (AFC). AFC quantizes or compresses CSI from 3 dimensions --- time, frequency and numerical values, and adapts the intensity of compression according to channel profile. This simple principle faces many practical challenges, e.g., a huge search space for adaption, estimation or prediction of the impact of compression on network throughput, and the coupling of different users in multi-user MIMO networks. AFC meets these challenges using a novel cross-layer adaptation metric, a metric extracted from 802.11 packet preambles, and uses it to guide the selection of compression intensity, so as to balance the tradeoff between overhead reduction and capacity loss (due to compression). We have implemented AFC on a software radio testbed. Our experiments show that AFC can outperform alternative approaches in a variety of radio environments.