Liu Liu

Siyuan Zhou, Liu Liu, Li Niu et al.

Object placement aims to place a foreground object over a background image with a suitable location and size. In this work, we treat object placement as a graph completion problem and propose a novel graph completion module (GCM). The background scene is represented by a graph with multiple nodes at different spatial locations with various receptive fields. The foreground object is encoded as a special node that should be inserted at a reasonable place in this graph. We also design a dual-path framework upon the structure of GCM to fully exploit annotated composite images. With extensive experiments on OPA dataset, our method proves to significantly outperform existing methods in generating plausible object placement without loss of diversity.

Rui Xie, Asad Ul Haq, Linsen Ma et al.

Recent studies have revealed that, during the inference on generative AI models such as transformer, the importance of different weights exhibits substantial context-dependent variations. This naturally manifests a promising potential of adaptively configuring weight quantization to improve the generative AI inference efficiency. Although configurable weight quantization can readily leverage the hardware support of variable-precision arithmetics in modern GPU and AI accelerators, little prior research has studied how one could exploit variable weight quantization to proportionally improve the AI model memory access speed and energy efficiency. Motivated by the rapidly maturing CXL ecosystem, this work develops a CXL-based design solution to fill this gap. The key is to allow CXL memory controllers play an active role in supporting and exploiting runtime configurable weight quantization. Using transformer as a representative generative AI model, we carried out experiments that well demonstrate the effectiveness of the proposed design solution.

Guangqian Yang, Yibing Zhan, Jinlong Li et al.

Recent studies show that Graph Neural Networks (GNNs) are vulnerable to adversarial attack, i.e., an imperceptible structure perturbation can fool GNNs to make wrong predictions. Some researches explore specific properties of clean graphs such as the feature smoothness to defense the attack, but the analysis of it has not been well-studied. In this paper, we analyze the adversarial attack on graphs from the perspective of feature smoothness which further contributes to an efficient new adversarial defensive algorithm for GNNs. We discover that the effect of the high-order graph structure is a smoother filter for processing graph structures. Intuitively, the high-order graph structure denotes the path number between nodes, where larger number indicates closer connection, so it naturally contributes to defense the adversarial perturbation. Further, we propose a novel algorithm that incorporates the high-order structural information into the graph structure learning. We perform experiments on three popular benchmark datasets, Cora, Citeseer and Polblogs. Extensive experiments demonstrate the effectiveness of our method for defending against graph adversarial attacks.

Ran Zhang, Xuanhua He, Li Xueheng et al.

The field of pan-sharpening has recently seen a trend towards increasingly large and complex models, often trained on single, specific satellite datasets. This one-dataset, one-model approach leads to high computational overhead and impractical deployment. More critically, it overlooks a core challenge: poor generalization from reduced-resolution (RR) training to real-world full-resolution (FR) data. In response to this issue, we challenge this paradigm. We introduce a multiple-in-one training strategy, where a single, compact model is trained simultaneously on three distinct satellite datasets (WV2, WV3, and GF2). Our experiments show the primary benefit of this unified strategy is a significant and universal boost in FR generalization (QNR) across all tested models, directly addressing this overlooked problem. This paradigm also inherently solves the one-model-per-dataset challenge, and we support it with a highly reproducible, dependency-free codebase for true usability. Finally, we propose PanTiny, a lightweight framework designed specifically for this new, robust paradigm. We demonstrate it achieves a superior performance-to-efficiency balance, proving that principled, simple and robust design is more effective than brute-force scaling in this practical setting. Our work advocates for a community-wide shift towards creating efficient, deployable, and truly generalizable models for pan-sharpening. The code is open-sourced at https://github.com/Zirconium233/PanTiny.

Wenyan Cong, Jianfu Zhang, Li Niu et al.

Image composition is an important operation in image processing, but the inconsistency between foreground and background significantly degrades the quality of composite image. Image harmonization, which aims to make the foreground compatible with the background, is a promising yet challenging task. However, the lack of high-quality public dataset for image harmonization, which significantly hinders the development of image harmonization techniques. Therefore, we contribute an image harmonization dataset iHarmony4 by generating synthesized composite images based on existing COCO (resp., Adobe5k, day2night) dataset, leading to our HCOCO (resp., HAdobe5k, Hday2night) sub-dataset. To enrich the diversity of our dataset, we also generate synthesized composite images based on our collected Flick images, leading to our HFlickr sub-dataset. The image harmonization dataset iHarmony4 is released at https://github.com/bcmi/Image_Harmonization_Datasets.

Zhaodong Chen, Yuying Quan, Zheng Qu et al.

Transformers are becoming the mainstream solutions for various tasks like NLP and Computer vision. Despite their success, the high complexity of the attention mechanism hinders them from being applied to latency-sensitive tasks. Tremendous efforts have been made to alleviate this problem, and many of them successfully reduce the asymptotic complexity to linear. Nevertheless, most of them fail to achieve practical speedup over the original full attention under moderate sequence lengths and are unfriendly to finetuning. In this paper, we present DFSS, an attention mechanism that dynamically prunes the full attention weight matrix to N:M fine-grained structured sparse pattern. We provide both theoretical and empirical evidence that demonstrates DFSS is a good approximation of the full attention mechanism. We propose a dedicated CUDA kernel design that completely eliminates the dynamic pruning overhead and achieves speedups under arbitrary sequence length. We evaluate the 1:2 and 2:4 sparsity under different configurations and achieve 1.27~ 1.89x speedups over the full-attention mechanism. It only takes a couple of finetuning epochs from the pretrained model to achieve on par accuracy with full attention mechanism on tasks from various domains under different sequence lengths from 384 to 4096.

Chao Chen, Yibing Zhan, Baosheng Yu et al.

Scene Graph Generation (SGG) aims to build a structured representation of a scene using objects and pairwise relationships, which benefits downstream tasks. However, current SGG methods usually suffer from sub-optimal scene graph generation because of the long-tailed distribution of training data. To address this problem, we propose Resistance Training using Prior Bias (RTPB) for the scene graph generation. Specifically, RTPB uses a distributed-based prior bias to improve models' detecting ability on less frequent relationships during training, thus improving the model generalizability on tail categories. In addition, to further explore the contextual information of objects and relationships, we design a contextual encoding backbone network, termed as Dual Transformer (DTrans). We perform extensive experiments on a very popular benchmark, VG150, to demonstrate the effectiveness of our method for the unbiased scene graph generation. In specific, our RTPB achieves an improvement of over 10% under the mean recall when applied to current SGG methods. Furthermore, DTrans with RTPB outperforms nearly all state-of-the-art methods with a large margin.

Liu Liu, Zheng Qu, Zhaodong Chen et al.

Transformers are the mainstream of NLP applications and are becoming increasingly popular in other domains such as Computer Vision. Despite the improvements in model quality, the enormous computation costs make Transformers difficult at deployment, especially when the sequence length is large in emerging applications. Processing attention mechanism as the essential component of Transformer is the bottleneck of execution due to the quadratic complexity. Prior art explores sparse patterns in attention to support long sequence modeling, but those pieces of work are on static or fixed patterns. We demonstrate that the sparse patterns are dynamic, depending on input sequences. Thus, we propose the Dynamic Sparse Attention (DSA) that can efficiently exploit the dynamic sparsity in the attention of Transformers. Compared with other methods, our approach can achieve better trade-offs between accuracy and model complexity. Moving forward, we identify challenges and provide solutions to implement DSA on existing hardware (GPUs) and specialized hardware in order to achieve practical speedup and efficiency improvements for Transformer execution.

Yang Lyu, Thien-Minh Nguyen, Liu Liu et al.

Although Simultaneous Localization and Mapping (SLAM) has been an active research topic for decades, current state-of-the-art methods still suffer from instability or inaccuracy due to feature insufficiency or its inherent estimation drift, in many civilian environments. To resolve these issues, we propose a navigation system combing the SLAM and prior-map-based localization. Specifically, we consider additional integration of line and plane features, which are ubiquitous and more structurally salient in civilian environments, into the SLAM to ensure feature sufficiency and localization robustness. More importantly, we incorporate general prior map information into the SLAM to restrain its drift and improve the accuracy. To avoid rigorous association between prior information and local observations, we parameterize the prior knowledge as low dimensional structural priors defined as relative distances/angles between different geometric primitives. The localization is formulated as a graph-based optimization problem that contains sliding-window-based variables and factors, including IMU, heterogeneous features, and structure priors. We also derive the analytical expressions of Jacobians of different factors to avoid the automatic differentiation overhead. To further alleviate the computation burden of incorporating structural prior factors, a selection mechanism is adopted based on the so-called information gain to incorporate only the most effective structure priors in the graph optimization. Finally, the proposed framework is extensively tested on synthetic data, public datasets, and, more importantly, on the real UAV flight data obtained from a building inspection task. The results show that the proposed scheme can effectively improve the accuracy and robustness of localization for autonomous robots in civilian applications.

Liu Liu, Hongdong Li, Haodong Yao et al.

Aligning two partially-overlapped 3D line reconstructions in Euclidean space is challenging, as we need to simultaneously solve correspondences and relative pose between line reconstructions. This paper proposes a neural network based method and it has three modules connected in sequence: (i) a Multilayer Perceptron (MLP) based network takes Pluecker representations of lines as inputs, to extract discriminative line-wise features and matchabilities (how likely each line is going to have a match), (ii) an Optimal Transport (OT) layer takes two-view line-wise features and matchabilities as inputs to estimate a 2D joint probability matrix, with each item describes the matchness of a line pair, and (iii) line pairs with Top-K matching probabilities are fed to a 2-line minimal solver in a RANSAC framework to estimate a six Degree-of-Freedom (6-DoF) rigid transformation. Experiments on both indoor and outdoor datasets show that the registration (rotation and translation) precision of our method outperforms baselines significantly.

Ziqiang Shi, Liu Liu, Zhongling Liu et al.

The detection of facial action units (AUs) has been studied as it has the competition due to the wide-ranging applications thereof. In this paper, we propose a novel framework for the AU detection from a single input image by grasping the \textbf{c}o-\textbf{o}ccurrence and \textbf{m}utual \textbf{ex}clusion (COMEX) as well as the intensity distribution among AUs. Our algorithm uses facial landmarks to detect the features of local AUs. The features are input to a bidirectional long short-term memory (BiLSTM) layer for learning the intensity distribution. Afterwards, the new AU feature continuously passed through a self-attention encoding layer and a continuous-state modern Hopfield layer for learning the COMEX relationships. Our experiments on the challenging BP4D and DISFA benchmarks without any external data or pre-trained models yield F1-scores of 63.7\% and 61.8\% respectively, which shows our proposed networks can lead to performance improvement in the AU detection task.

Xibin Song, Yuchao Dai, Dingfu Zhou et al.

Despite the remarkable progresses made in deep-learning based depth map super-resolution (DSR), how to tackle real-world degradation in low-resolution (LR) depth maps remains a major challenge. Existing DSR model is generally trained and tested on synthetic dataset, which is very different from what would get from a real depth sensor. In this paper, we argue that DSR models trained under this setting are restrictive and not effective in dealing with real-world DSR tasks. We make two contributions in tackling real-world degradation of different depth sensors. First, we propose to classify the generation of LR depth maps into two types: non-linear downsampling with noise and interval downsampling, for which DSR models are learned correspondingly. Second, we propose a new framework for real-world DSR, which consists of four modules : 1) An iterative residual learning module with deep supervision to learn effective high-frequency components of depth maps in a coarse-to-fine manner; 2) A channel attention strategy to enhance channels with abundant high-frequency components; 3) A multi-stage fusion module to effectively re-exploit the results in the coarse-to-fine process; and 4) A depth refinement module to improve the depth map by TGV regularization and input loss. Extensive experiments on benchmarking datasets demonstrate the superiority of our method over current state-of-the-art DSR methods.

Fei Sun, Minghai Qin, Tianyun Zhang et al.

Neural network models are widely used in solving many challenging problems, such as computer vision, personalized recommendation, and natural language processing. Those models are very computationally intensive and reach the hardware limit of the existing server and IoT devices. Thus, finding better model architectures with much less amount of computation while maximally preserving the accuracy is a popular research topic. Among various mechanisms that aim to reduce the computation complexity, identifying the zero values in the model weights and in the activations to avoid computing them is a promising direction. In this paper, we summarize the current status of the research on the computation of sparse neural networks, from the perspective of the sparse algorithms, the software frameworks, and the hardware accelerations. We observe that the search for the sparse structure can be a general methodology for high-quality model explorations, in addition to a strategy for high-efficiency model execution. We discuss the model accuracy influenced by the number of weight parameters and the structure of the model. The corresponding models are called to be located in the weight dominated and structure dominated regions, respectively. We show that for practically complicated problems, it is more beneficial to search large and sparse models in the weight dominated region. In order to achieve the goal, new approaches are required to search for proper sparse structures, and new sparse training hardware needs to be developed to facilitate fast iterations of sparse models.

Yujiao Shi, Xin Yu, Liu Liu et al.

This paper addresses the problem of cross-view image geo-localization, where the geographic location of a ground-level street-view query image is estimated by matching it against a large scale aerial map (e.g., a high-resolution satellite image). State-of-the-art deep-learning based methods tackle this problem as deep metric learning which aims to learn global feature representations of the scene seen by the two different views. Despite promising results are obtained by such deep metric learning methods, they, however, fail to exploit a crucial cue relevant for localization, namely, the spatial layout of local features. Moreover, little attention is paid to the obvious domain gap (between aerial view and ground view) in the context of cross-view localization. This paper proposes a novel Cross-View Feature Transport (CVFT) technique to explicitly establish cross-view domain transfer that facilitates feature alignment between ground and aerial images. Specifically, we implement the CVFT as network layers, which transports features from one domain to the other, leading to more meaningful feature similarity comparison. Our model is differentiable and can be learned end-to-end. Experiments on large-scale datasets have demonstrated that our method has remarkably boosted the state-of-the-art cross-view localization performance, e.g., on the CVUSA dataset, with significant improvements for top-1 recall from 40.79% to 61.43%, and for top-10 from 76.36% to 90.49%. We expect the key insight of the paper (i.e., explicitly handling domain difference via domain transport) will prove to be useful for other similar problems in computer vision as well.

Liu Liu, Lei Deng, Xing Hu et al.

We propose to execute deep neural networks (DNNs) with dynamic and sparse graph (DSG) structure for compressive memory and accelerative execution during both training and inference. The great success of DNNs motivates the pursuing of lightweight models for the deployment onto embedded devices. However, most of the previous studies optimize for inference while neglect training or even complicate it. Training is far more intractable, since (i) the neurons dominate the memory cost rather than the weights in inference; (ii) the dynamic activation makes previous sparse acceleration via one-off optimization on fixed weight invalid; (iii) batch normalization (BN) is critical for maintaining accuracy while its activation reorganization damages the sparsity. To address these issues, DSG activates only a small amount of neurons with high selectivity at each iteration via a dimension-reduction search (DRS) and obtains the BN compatibility via a double-mask selection (DMS). Experiments show significant memory saving (1.7-4.5x) and operation reduction (2.3-4.4x) with little accuracy loss on various benchmarks.

Liu Liu, Yanyao Shen, Tianyang Li et al.

We provide a novel -- and to the best of our knowledge, the first -- algorithm for high dimensional sparse regression with constant fraction of corruptions in explanatory and/or response variables. Our algorithm recovers the true sparse parameters with sub-linear sample complexity, in the presence of a constant fraction of arbitrary corruptions. Our main contribution is a robust variant of Iterative Hard Thresholding. Using this, we provide accurate estimators: when the covariance matrix in sparse regression is identity, our error guarantee is near information-theoretically optimal. We then deal with robust sparse regression with unknown structured covariance matrix. We propose a filtering algorithm which consists of a novel randomized outlier removal technique for robust sparse mean estimation that may be of interest in its own right: the filtering algorithm is flexible enough to deal with unknown covariance. Also, it is orderwise more efficient computationally than the ellipsoid algorithm. Using sub-linear sample complexity, our algorithm achieves the best known (and first) error guarantee. We demonstrate the effectiveness on large-scale sparse regression problems with arbitrary corruptions.

Shuang Wu, Guoqi Li, Lei Deng et al.

Batch Normalization (BN) has been proven to be quite effective at accelerating and improving the training of deep neural networks (DNNs). However, BN brings additional computation, consumes more memory and generally slows down the training process by a large margin, which aggravates the training effort. Furthermore, the nonlinear square and root operations in BN also impede the low bit-width quantization techniques, which draws much attention in deep learning hardware community. In this work, we propose an L1-norm BN (L1BN) with only linear operations in both the forward and the backward propagations during training. L1BN is shown to be approximately equivalent to the original L2-norm BN (L2BN) by multiplying a scaling factor. Experiments on various convolutional neural networks (CNNs) and generative adversarial networks (GANs) reveal that L1BN maintains almost the same accuracies and convergence rates compared to L2BN but with higher computational efficiency. On FPGA platform, the proposed signum and absolute operations in L1BN can achieve 1.5$\times$ speedup and save 50\% power consumption, compared with the original costly square and root operations, respectively. This hardware-friendly normalization method not only surpasses L2BN in speed, but also simplify the hardware design of ASIC accelerators with higher energy efficiency. Last but not the least, L1BN promises a fully quantized training of DNNs, which is crucial to future adaptive terminal devices.

Liu Liu, Shaoshan Liu, Zhe Zhang et al.

Enabling full robotic workloads with diverse behaviors on mobile systems with stringent resource and energy constraints remains a challenge. In recent years, attempts have been made to deploy single-accelerator-based computing platforms (such as GPU, DSP, or FPGA) to address this challenge, but with little success. The core problem is two-fold: firstly, different robotic tasks require different accelerators, and secondly, managing multiple accelerators simultaneously is overwhelming for developers. In this paper, we propose PIRT, the first robotic runtime framework to efficiently manage dynamic task executions on mobile systems with multiple accelerators as well as on the cloud to achieve better performance and energy savings. With PIRT, we enable a robot to simultaneously perform autonomous navigation with 25 FPS of localization, obstacle detection with 3 FPS, route planning, large map generation, and scene understanding, traveling at a max speed of 5 miles per hour, all within an 11W computing power envelope.

Ali Taalimi, Alireza Rahimpour, Liu Liu et al.

Nowadays, distributed smart cameras are deployed for a wide set of tasks in several application scenarios, ranging from object recognition, image retrieval, and forensic applications. Due to limited bandwidth in distributed systems, efficient coding of local visual features has in fact been an active topic of research. In this paper, we propose a novel approach to obtain a compact representation of high-dimensional visual data using sensor fusion techniques. We convert the problem of visual analysis in resource-limited scenarios to a multi-view representation learning, and we show that the key to finding properly compressed representation is to exploit the position of cameras with respect to each other as a norm-based regularization in the particular signal representation of sparse coding. Learning the representation of each camera is viewed as an individual task and a multi-task learning with joint sparsity for all nodes is employed. The proposed representation learning scheme is referred to as the multi-view task-driven learning for visual sensor network (MT-VSN). We demonstrate that MT-VSN outperforms state-of-the-art in various surveillance recognition tasks.

Tianyang Li, Liu Liu, Anastasios Kyrillidis et al.

We present a novel method for frequentist statistical inference in $M$-estimation problems, based on stochastic gradient descent (SGD) with a fixed step size: we demonstrate that the average of such SGD sequences can be used for statistical inference, after proper scaling. An intuitive analysis using the Ornstein-Uhlenbeck process suggests that such averages are asymptotically normal. From a practical perspective, our SGD-based inference procedure is a first order method, and is well-suited for large scale problems. To show its merits, we apply it to both synthetic and real datasets, and demonstrate that its accuracy is comparable to classical statistical methods, while requiring potentially far less computation.

Maohua Zhu, Liu Liu, Chao Wang et al.

Designing and implementing efficient, provably correct parallel neural network processing is challenging. Existing high-level parallel abstractions like MapReduce are insufficiently expressive while low-level tools like MPI and Pthreads leave ML experts repeatedly solving the same design challenges. However, the diversity and large-scale data size have posed a significant challenge to construct a flexible and high-performance implementation of deep learning neural networks. To improve the performance and maintain the scalability, we present CNNLab, a novel deep learning framework using GPU and FPGA-based accelerators. CNNLab provides a uniform programming model to users so that the hardware implementation and the scheduling are invisible to the programmers. At runtime, CNNLab leverages the trade-offs between GPU and FPGA before offloading the tasks to the accelerators. Experimental results on the state-of-the-art Nvidia K40 GPU and Altera DE5 FPGA board demonstrate that the CNNLab can provide a universal framework with efficient support for diverse applications without increasing the burden of the programmers. Moreover, we analyze the detailed quantitative performance, throughput, power, energy, and performance density for both approaches. Experimental results leverage the trade-offs between GPU and FPGA and provide useful practical experiences for the deep learning research community.