Video-based Point Cloud Compression (V-PCC) is an emerging standard for encoding dynamic point cloud data. With V-PCC, point cloud data is segmented, projected, and packed on to 2D video frames, which can be compressed using existing video coding standards such as H.264, H.265 and AV1. This makes it possible to support point cloud streaming via reliable video transmission systems. On the other hand, despite recent advances, many issues still remain and prevent V-PCC from being used in low-latency point cloud streaming. For instance, point cloud registration and patch generation can take a long time. In this paper, we focus on one unique problem in V-PCC: bit allocation among different sub-streams - the geometry sub-stream and the attribute (color) sub-stream - with the goal of improving the visual quality of point clouds under the target bitrate. Existing approaches either do not fully utilize the available bandwidth or can take a long time to run, which cannot be used in scenarios that require low-latency. To this end, we propose a lightweight, frequency-domain-based profiling method for transforming the dynamic point cloud data into a one-dimension vector. By using two single-layer linear regression models, we can estimate the compressed bitrate for geometry data and color information. This allows us to perform bit allocation between the geometry map and the attribute map with simple calculations. Evaluation results show that compared to the baseline approach, our method can achieve better visual qualities with smaller encoded segment sizes under the target bitrate.
论文检索
输入标题、作者或关键词，从 1,014 篇学术成果中精准定位
Backdoor attack is a major threat to deep learning systems in safety-critical scenarios, which aims to trigger misbehavior of neural network models under attacker-controlled conditions. However, most backdoor attacks have to modify the neural network models through training with poisoned data and/or direct model editing, which leads to a common but false belief that backdoor attack can be easily avoided by properly protecting the model. In this paper, we show that backdoor attacks can be achieved without any model modification. Instead of injecting backdoor logic into the training data or the model, we propose to place a carefully-designed patch (namely backdoor patch) in front of the camera, which is fed into the model together with the input images. The patch can be trained to behave normally at most of the time, while producing wrong prediction when the input image contains an attacker-controlled trigger object. Our main techniques include an effective training method to generate the backdoor patch and a digital-physical transformation modeling method to enhance the feasibility of the patch in real deployments. Extensive experiments show that PatchBackdoor can be applied to common deep learning models (VGG, MobileNet, ResNet) with an attack success rate of 93% to 99% on classification tasks. Moreover, we implement PatchBackdoor in real-world scenarios and show that the attack is still threatening.
Online 6-DoF view touring has become increasingly popular due to hardware advances and the recent pandemic. One way for content creators to support many 6-DoF clients is by transmitting 3D content to them, which leads to content leakage. Another way for content creators is to render and stream novel views for 6-DoF clients, which incurs staggering computational and networking workloads. In this paper, we develop a blind streaming system that leverages cloud service providers between content creators and 6-DoF clients. Our system has two core design objectives: (i) to generate high-quality novel views for 6-DoF clients without retrieving 3D content from content creators, (ii) to support many 6-DoF clients without overloading the content creators. We achieve these two goals in the following steps. First, we design a source view request/response interface between cloud service providers and content creators for efficient communications. Second, we design novel view optimization algorithms for cloud service providers to intelligently select the minimal set of source views while considering the workload of content creators. Third, we employ scalable client side view synthesis for 6-DoF clients with heterogeneous device capabilities and personalized 6-DoF client poses and preferences. Our evaluation results demonstrate the merits of our solution, compared to the state-of-the-arts, our system: (i) improves synthesized novel views by 2.27 dB in PSNR and 12 in VMAF on average and (ii) reduces the bandwidth consumption by 94% on average. In fact, our solution approaches the performance of an unrealistic optimal solution with unlimited source views, achieving performance gaps as small as 0.75 dB in PSNR and 3.8 in VMAF.
With the rapid advances of deep learning and the commercialization of high-definition cameras in mobile and embedded devices, the demands from latency-critical applications such as AR and XR for high-quality video analytics (HVA) are soaring. By the nature of HVA aiming at enabling detailed analytics even for small objects, its on-device implementation is suffering from thermal and battery issues, which makes offloaded HVA an attractive solution. This work provides unique observations on the tradeoff pertaining to offloaded HVA: the frame encoding time, the frame transmission time, and the HVA accuracy. Our observations pose a fundamental question: given a latency budget, how to choose the encoding option that properly combines between the encoding time and the transmission time to maximize the HVA accuracy. To answer this question, we propose an offloaded HVA system, ENTRO, which exploits this tradeoff in real-time to maximize the HVA accuracy under the latency budget. Our extensive evaluations with ENTRO implemented on Nvidia AGX Xavier and Samsung Galaxy S20 Ultra over WiFi networks show 8.8× improvement in latency without accuracy loss compared to DDS, the state-of-the-art offloaded video analytics. Our evaluation over commercial 5G and LTE networks also indicates that ENTRO flexibly adapts its encoding option under the tradeoff and enables the latency-bounded HVA with 4K frames.
Recently, the efficient deployment and acceleration of powerful vision transformers (ViTs) on resource-limited edge devices for providing multimedia services have become attractive tasks. Although early exiting is a feasible solution for accelerating inference, most works focus on convolutional neural networks (CNNs) and transformer models in natural language processing (NLP). Moreover, the direct application of early exiting methods to ViTs may result in substantial performance degradation. To tackle this challenge, we systematically investigate the efficacy of early exiting in ViTs and point out that the insufficient feature representations in shallow internal classifiers and the limited ability to capture target semantic information in deep internal classifiers restrict the performance of these methods. We then propose an early exiting framework for general ViTs termed LGViT, which incorporates heterogeneous exiting heads, namely, local perception head and global aggregation head, to achieve an efficiency-accuracy trade-off. In particular, we develop a novel two-stage training scheme, including end-to-end training and self-distillation with the backbone frozen to generate early exiting ViTs, which facilitates the fusion of global and local information extracted by the two types of heads. We conduct extensive experiments using three popular ViT backbones on three vision datasets. Results demonstrate that our LGViT can achieve competitive performance with approximately 1.8 × speed-up.
Recent years have witnessed the profound influence of AI technologies on computer gaming. While grandmaster-level AI robots have largely come true for complex games based on heavy back-end support, in practice many game developers crave for participant AI robots (PARs) that behave like average-level humans with inexpensive infrastructures. Unfortunately, to date there has not been a satisfactory solution that registers large-scale use. In this work, we attempt to develop practical PARs (dubbed ParliRobo) showing acceptably humanoid behaviors with well affordable infrastructures under a challenging scenario-a 3D-FPS (first-person shooter) mobile MMOG with real-time interaction requirements. Based on comprehensive real-world explorations, we eventually enable our attempt through a novel ?transform and polish" methodology. It achieves ultralight implementations of the core system components by non-intuitive yet principled approaches, and meanwhile carefully fixes the probable side effect incurred on user perceptions. Evaluation results from large-scale deployment indicate the close resemblance (96% on average) in biofidelity metrics between ParliRobo and human players; moreover, in 73% mini Turing tests ParliRobo cannot be distinguished from human players.
Neural Radiance Fields (NeRFs) have revolutionized 3D scene synthesis. Voxel grids are commonly employed to enhance training or rendering speed, but they entail additional storage requirements. The large model size and high computational and memory demands impede their progress on resource-constrained devices, e.g., Microcontroller Units (MCUs). Besides, there is currently no NeRF rendering framework available on MCU devices. In this paper, we propose a NeRF method named MCUNeRF for 3D scene synthesis on MCU devices. The proposed MCUNeRF compresses voxel grids via a hybrid quantization algorithm merging learned step-size quantization (LSQ) and optimized product quantization (OPQ). To further reduce the model storage, we also propose a codebook-sharing method that renders multiple objects with a single quantization codebook. Then we implement a NeRF-based rendering framework for MCU devices, which leverages a low-bit neural network computation framework, i.e. CMSIS-NN, to accelerate the rendering progress. Extensive experiments on four datasets such as Synthetic-NeRF demonstrate that our proposed method could compress model data by 20-40 times with comparable rendering quality, which enables NeRF-based scene rendering on MCU devices with only 1M SRAM.
A learning-based YOGA (Yet Another Geometry-based Point Cloud Compressor) is proposed. It is flexible, allowing for the separable lossy compression of geometry and color attributes, and variable-rate coding using a single neural model; it is high-efficiency, significantly outperforming the latest G-PCC standard quantitatively and qualitatively, e.g., 25% BD-BR gains using PCQM (Point Cloud Quality Metric) as the distortion assessment, and it is lightweight, e.g., similar runtime as the G-PCC codec, owing to the use of sparse convolution and parallel entropy coding. To this end, YOGA adopts a unified end-to-end learning-based backbone for separate geometry and attribute compression. The backbone uses a two-layer structure, where the downscaled thumbnail point cloud is encoded using G-PCC at the base layer, and upon G-PCC compressed priors, multiscale sparse convolutions are stacked at the enhancement layer to effectively characterize spatial correlations to compactly represent the full-resolution sample. In addition, YOGA integrates the adaptive quantization and entropy model group to enable variable-rate control, as well as adaptive filters for better quality restoration.
Single image super-resolution (SR) is an important low-level vision task, and the dynamic SR trading off performance and efficiency are increasingly in demand. The existing dynamic SR methods are divided into two classes: the structured pruning and non-structured compressing methods. The former removes redundant structures in the network, which often leads to significant performance degradation, and the latter searches for extremely sparse parameter masks, achieving promising performance, but they are not deployable in hardware platforms with irregular memory access. In order to solve the mentioned problems, we propose Hardware-friendly Scalable SR (HSSR) with progressively structured sparsity. The superiority of our method is that with only a single scalable model it covers multiple SR models with different sizes, without extra retraining or post-processing. HSSR contains the forward and backward processing. In the forward process, we gradually shrink the SR networks with structured iterative sparsity where grouping convolution together with knowledge distillation is conducted to reduce the amount of SR parameters and the computational complexity while keeping the performance, and in the backward process, we gradually expand the compressed SR networks with structured iterative recovery. Comprehensive experiments on benchmark datasets show that HSSR is perfectly compatible with common convolution baselines. Compared with the Slimmable method, our model is superior in performance, flops, and model size. Experimental results demonstrate that HSSR achieves significant compression, saving up to 1500K parameters and 100 GFlops calculation compared to the original model in real-world applications.
While large deep neural networks excel at general video analytics tasks, the significant demand on computing capacity makes them infeasible for real-time inference on resource-constrained end cameras. In this paper, we propose an edge-assisted framework that continuously updates the lightweight model deployed on the end cameras to achieve accurate predictions in adverse environments. This framework consists of three modules, namely, a key frame extractor, a trigger controller, and a retraining manager. The low-cost key frame extractor obtains frames that can best represent the current environment. Those frames are then transmitted and buffered as the retraining data for model update at the edge server. Once the trigger controller detects a significant accuracy drop in the selected frames, the retraining manager outputs the optimal retraining configuration balancing the accuracy and time cost. We prototype our system on two end devices of different computing capacities with one edge server. The results demonstrate that our approach significantly improves accuracy across all tested adverse environment scenarios (up to 24%) and reduces more than 50% of the retraining time compared to existing benchmarks.
Motion estimation (ME) is one of the most critical tools in video coding and consumes the majority of the encoding complexity. Three types of ME are utilized in the latest video coding standards, namely integer, fractional, and affine MEs. They are implemented as three searches for the integer motion vector (IMV), fractional motion vector (FMV), and control point motion vectors (CPMVs). Many algorithms were proposed to reduce the complexity for them individually, but the overall overhead of three searches is still challenging for hardware implementations. Therefore, we propose a hardware-efficient Unified Motion Estimation (UME) to derive three types of MVs with only one search. An IME with sub-block refinement is performed to collect extra motion information while searching for the IMV. The FMV and CPMVs are then derived from the collected information using a mixed error surface and an overdetermined system. Compared to the default ME algorithms in VVC, the time cost for ME is reduced by 41.63% with a coding loss of only 0.87% under LDB configuration. For hardware implementations, the minimum required resources and corresponding latency are significantly reduced by 75.35% and 69.17%, respectively.
Over the past years, augmented reality (AR) based on mobile phones has gained great attention. When multiple phones are used in AR applications, collaborative simultaneous localization and mapping (SLAM) is considered one of the enabling technologies, i.e., multiple mobile phones complete the localization and mapping through collaboration. However, the state-of-the-art collaborative SLAM systems not only suffer from the delays introduced by a high-complexity graph optimization problem, but also may exhibit varying levels of accuracy across dissimilar environments or different types of mobile devices. In this paper, we propose a scalable and robust collaborative SLAM system, point-line-based Collaborative SLAM (ColSLAM). Technically, ColSLAM includes two innovative features that help achieve satisfactory scalability and robustness. First, a mapping cacher (MC) is designed for each agent on the server, which uses global keyframes to detect loop closures, updates the cached local map, and quickly responds to the agent's pose drifts. With MC, each agent's local pose is corrected using global knowledge in real-time. Secondly, to improve the robustness performance, ColSLAM employs point-line-fusion-based Visual Inertial Odometry (VIO), point-line-fusion-based NetVLAD loop detection, and an enhanced geometric verification and relative pose calculation method called PNPL. Empirical evaluations based on the EuRoc dataset and real degenerate environments demonstrate that ColSLAM outperforms the existing collaborative SLAM systems in terms of accuracy, robustness, and scalability.
NIF: A Fast Implicit Image Compression with Bottleneck Layers and Modulated Sinusoidal Activations
PDF ↗In Implicit Neural Representations (INRs) a discrete signal is parameterized by a neural network that maps coordinates to the signal samples. INRs were successfully employed for encoding and compression, but such approaches are in their early stage and are still overcome by traditional codecs and autoencoders. Despite this, they have recently gained the attention of the research community due to their promising results as novel representation strategies for encoding visual content. In this paper, we propose Neural Imaging Format (NIF), an open-source INR-based image compression codec which takes advantage of a novel neural architecture which consists of two modules: a Genesis network, for mapping coordinates to pixels through bottleneck layers with sinusoidal activation units, and a Modulation network, for varying the period of the sinusoidal activations. Additionally, a final weights quantization step leads to an improvement in the compression ratio. Our proposal (NIF) consistently outperforms state-of-art INR-based compressors in terms of PSNR, by achieving comparable or better results with an outstanding up to X26 encoding speed. We also show that NIF reduces the gap between INR-based methods with respect to traditional approaches. Interestingly, our approach outperforms established codecs such as JPEG and WebP when one encodes high-resolution images at low-bitrate regimes. Extensive experiments on different datasets, a visual comparison, and an ablation study, prove the validity of the proposed approach.
3D object detection plays a pivotal role in many applications, most notably autonomous driving and robotics. These applications are commonly deployed on edge devices to promptly interact with the environment, and often require near real-time response. With limited computation power, it is challenging to execute 3D detection on the edge using highly complex neural networks. Common approaches such as offloading to the cloud induce significant latency overheads due to the large amount of point cloud data during transmission. To resolve the tension between wimpy edge devices and compute-intensive inference workloads, we explore the possibility of empowering fast 2D detection to extrapolate 3D bounding boxes. To this end, we present Moby, a novel system that demonstrates the feasibility and potential of our approach. We design a transformation pipeline for Moby that generates 3D bounding boxes efficiently and accurately based on 2D detection results without running 3D detectors. Further, we devise a frame offloading scheduler that decides when to launch the 3D detector judiciously in the cloud to avoid the errors from accumulating. Extensive evaluations on NVIDIA Jetson TX2 with real-world autonomous driving datasets demonstrate that Moby offers up to 91.9% latency improvement with modest accuracy loss over state of the art.
With the growth of deep neural networks (DNN), the number of DNN parameters has drastically increased. This makes DNN models hard to be deployed on resource-limited embedded systems. To alleviate this problem, dynamic pruning methods have emerged, which try to find diverse sparsity patterns during training by utilizing Straight-Through-Estimator (STE) to approximate gradients of pruned weights. STE can help the pruned weights revive in the process of finding dynamic sparsity patterns. However, using these coarse gradients causes training instability and performance degradation owing to the unreliable gradient signal of the STE approximation. In this work, to tackle this issue, we introduce refined gradients to update the pruned weights by forming dual forwarding paths from two sets (pruned and unpruned) of weights. We propose a novel Dynamic Collective Intelligence Learning (DCIL) which makes use of the learning synergy between the collective intelligence of both weight sets. We verify the usefulness of the refined gradients by showing enhancements in the training stability and the model performance on the CIFAR and ImageNet datasets. DCIL outperforms various previously proposed pruning schemes including other dynamic pruning methods with enhanced stability during training. The code is provided in Github.
One big challenge to federated learning is the non-IID data distribution caused by imbalanced classes. Existing federated learning approaches tend to bias towards classes containing a larger number of samples during local updates, which causes unwanted drift in the local classifiers. To address this issue, we propose a classifier debiased federated learning framework named FedCD for non-IID data. We introduce a novel hierarchical prototype contrastive learning strategy to learn fine-grained prototypes for each class. The prototypes characterize the sample distribution within each class, which helps align the features learned in the representation layer of every client's local model. At the representation layer, we use fine-grained prototypes to rebalance the class distribution on each client and rectify the classification layer of each local model. To alleviate the bias of the classification layer of the local models, we incorporate a global information distillation method to enable the local classifier to learn decoupled global classification information. We also adaptively aggregate the class-level classifiers based on their quality to reduce the impact of unreliable classes in each aggregated classifier. This mitigates the impact of client-side classifier bias on the global classifier. Comprehensive experiments conducted on various datasets show that our method, FedCD, effectively corrects classifier bias and outperforms state-of-the-art federated learning methods.
The graph has recently enabled substantial advances in big data analysis. As graphs are increasing from billions to trillions, efficient graph processing requires large-scale distributed clusters, which have up to thousands of nodes. For big data applications of which the computation is relatively simple, while the communication, especially for imbalanced communication is the bottleneck on distributed clusters, where huge numbers of small messages are transferred through 2D-topology networks. Graph partitioning is the dominant factor to affect the performance of large-scale distributed graph processing. Current graph partitioning policies have paid extensive attention to the utilization of the power law of big graphs but failed to exploit the advanced architectural benefits of 2D topology. To address such a problem, this paper presents GraphMedia, a communication-balanced graph partitioning for distributed search at scale. The key idea of GraphMedia is a communication-balanced partitioning to balance communication based on hardware/software co-design, in which the power law of graphs would be explored to average communication among nodes, and communication would be balanced between row and column by leveraging advanced 2D-topology knowledge. We use both benchmarks and real-world graphs to validate GraphMedia. Specially, GraphMedia-based Graph500 tests on the Tianhe supercomputer are superior to the fastest systems in the latest Graph500 lists (June 2022). We finally apply GraphMedia to real-world graphs for online graph media access, which outperforms the state-of-the-art graph partitioning and graph system by orders of magnitude.
In recent years, many perceptual authentication hashing schemes have been proposed, especially for image content authentication. However, most of the schemes directly use the dataset of image processing during model training and evaluation, which is actually unreasonable due to the task difference. In this paper, we first propose a specialized dataset for perceptual authentication hashing of images (PAHI), and the image content-preserving manipulations used in this dataset are richer and more in line with realistic scenarios. Then, in order to achieve satisfactory perceptual robustness and discrimination capability of PAHI, we exploit the continuous neural architecture search (NAS) on the channel number and stack depth of the ConvNeXt architecture, and obtain two PAHI architectures i.e., NASRes and NASCoNt. The former has better overall performance, while the latter is better for some special manipulations such as image cropping and background overlap. Experimental results demonstrate that our architectures both can achieve competitive results compared with SOTA schemes, and the AUC areas are increased by 1.6 (NASCoNt) and 1.7 (NASRes), respectively.
Massive numbers of new images are uploaded to the internet every day. However, existing cross-modal retrieval (CMR) approaches struggle to accommodate this continuously growing data. The prevalent practice involves periodically retraining or fine-tuning a new model based on the accumulated data, which in turn invalidates billions of indexed features extracted by the previous model and incurs another substantial computational cost to extract new features for the entire data archive. Is it possible to develop a retrieval model that effectively captures the knowledge of upcoming sessions while preserving the discriminative power of features extracted in previous sessions? In this paper, we propose an online continual learning setup, OC-CMR, to formalize the data-incremental growth challenge faced by cross-modal retrieval systems. It consists of two key settings: 1) Similar to the real-world scenarios, the streaming multi-modal data arrives once per session; 2) Consider the computational costs, each instance of archived data has its feature extracted only once and by its corresponding model in its session. Based on our OC-CMR, we perform in-depth evaluations of state-of-the-art cross-modal retrieval methods and observe that they suffer from representational shift and collapse due to the catastrophic forgetting. To address this issue, we propose the Continual Cross-Modal Retrieval (C2MR) approach, which learns a shared common space not only across modalities but also sessions and maintains relationships between samples from distinct sessions via cross-modal relational coherence and semantic representation coordination. We construct two new benchmarks by adapting MS-COCO and Flickr30K datasets to the OC-CMR setting, providing a more challenging evaluation framework for CMR tasks. Experimental results demonstrate that our method effectively alleviates forgetting and significantly outperforms combinations of previous arts in cross-modal retrieval and continual learning.
Temporal Knowledge Graph Embedding (TKGE) is a crucial technique for performing Temporal Knowledge Graph Completion (TKGC). The effectiveness of TKGE largely depends on the ability to model intrinsic relation patterns. However, as we know, most existing TKGE models usually embed KGs into a single geometric space such as Euclidean, hyperbolic or hyperspherical space to maintain their specific geometric structures (e.g., chain, hierarchy, and ring structures). None of the existing methods can simultaneously model relation patterns of chain, hierarchy, ring structures, and relation mapping properties. This paper constructs a hybrid interaction TKGE model HyIE, which learns spatial structures interactively between the Euclidean, hyperbolic and hyperspherical spaces. HyIE performs two Householder transformations of head and tail entities parameterized by relations in a high-dimensional mixed vector space. The curvature of hyperbolic and hyperspherical spaces depends on the product of both relation and temporal. The core of HyIE lies in implementing transformations and interactions of vectors in Euclidean, hyperbolic and hyperspherical spaces, and Household transformation of head and tail entities. Theoretically, HyIE can model crucial relation patterns and mapping properties simultaneously. Experimental results on five temporal knowledge graph benchmarks show that our HyIE achieves state-of-the-art performance.