Yi Wang

Kunchang Li, Yali Wang, Yinan He et al.

Learning discriminative spatiotemporal representation is the key problem of video understanding. Recently, Vision Transformers (ViTs) have shown their power in learning long-term video dependency with self-attention. Unfortunately, they exhibit limitations in tackling local video redundancy, due to the blind global comparison among tokens. UniFormer has successfully alleviated this issue, by unifying convolution and self-attention as a relation aggregator in the transformer format. However, this model has to require a tiresome and complicated image-pretraining phrase, before being finetuned on videos. This blocks its wide usage in practice. On the contrary, open-sourced ViTs are readily available and well-pretrained with rich image supervision. Based on these observations, we propose a generic paradigm to build a powerful family of video networks, by arming the pretrained ViTs with efficient UniFormer designs. We call this family UniFormerV2, since it inherits the concise style of the UniFormer block. But it contains brand-new local and global relation aggregators, which allow for preferable accuracy-computation balance by seamlessly integrating advantages from both ViTs and UniFormer. Without any bells and whistles, our UniFormerV2 gets the state-of-the-art recognition performance on 8 popular video benchmarks, including scene-related Kinetics-400/600/700 and Moments in Time, temporal-related Something-Something V1/V2, untrimmed ActivityNet and HACS. In particular, it is the first model to achieve 90% top-1 accuracy on Kinetics-400, to our best knowledge. Code will be available at https://github.com/OpenGVLab/UniFormerV2.

Keqiang Sun, Junting Pan, Yuying Ge et al. · pku

While recent advancements in vision-language models have had a transformative impact on multi-modal comprehension, the extent to which these models possess the ability to comprehend generated images remains uncertain. Synthetic images, in comparison to real data, encompass a higher level of diversity in terms of both content and style, thereby presenting significant challenges for the models to fully grasp. In light of this challenge, we introduce a comprehensive dataset, referred to as JourneyDB, that caters to the domain of generative images within the context of multi-modal visual understanding. Our meticulously curated dataset comprises 4 million distinct and high-quality generated images, each paired with the corresponding text prompts that were employed in their creation. Furthermore, we additionally introduce an external subset with results of another 22 text-to-image generative models, which makes JourneyDB a comprehensive benchmark for evaluating the comprehension of generated images. On our dataset, we have devised four benchmarks to assess the performance of generated image comprehension in relation to both content and style interpretation. These benchmarks encompass prompt inversion, style retrieval, image captioning, and visual question answering. Lastly, we evaluate the performance of state-of-the-art multi-modal models when applied to the JourneyDB dataset, providing a comprehensive analysis of their strengths and limitations in comprehending generated content. We anticipate that the proposed dataset and benchmarks will facilitate further research in the field of generative content understanding. The dataset is publicly available at https://journeydb.github.io.

Wenyang Liu, Yi Wang, Kejun Wu et al.

File fragment classification (FFC) on small chunks of memory is essential in memory forensics and Internet security. Existing methods mainly treat file fragments as 1d byte signals and utilize the captured inter-byte features for classification, while the bit information within bytes, i.e., intra-byte information, is seldom considered. This is inherently inapt for classifying variable-length coding files whose symbols are represented as the variable number of bits. Conversely, we propose Byte2Image, a novel data augmentation technique, to introduce the neglected intra-byte information into file fragments and re-treat them as 2d gray-scale images, which allows us to capture both inter-byte and intra-byte correlations simultaneously through powerful convolutional neural networks (CNNs). Specifically, to convert file fragments to 2d images, we employ a sliding byte window to expose the neglected intra-byte information and stack their n-gram features row by row. We further propose a byte sequence \& image fusion network as a classifier, which can jointly model the raw 1d byte sequence and the converted 2d image to perform FFC. Experiments on FFT-75 dataset validate that our proposed method can achieve notable accuracy improvements over state-of-the-art methods in nearly all scenarios. The code will be released at https://github.com/wenyang001/Byte2Image.

Gong Chen, Yanan Zhao, Yi Wang et al.

Recently, synthetic aperture radar (SAR) image change detection has become an interesting yet challenging direction due to the presence of speckle noise. Although both traditional and modern learning-driven methods attempted to overcome this challenge, deep convolutional neural networks (DCNNs)-based methods are still hindered by the lack of interpretability and the requirement of large computation power. To overcome this drawback, wavelet scattering network (WSN) and Fourier scattering network (FSN) are proposed. Combining respective merits of WSN and FSN, we propose Stockwell scattering network (SSN) based on Stockwell transform which is widely applied against noisy signals and shows advantageous characteristics in speckle reduction. The proposed SSN provides noise-resilient feature representation and obtains state-of-art performance in SAR image change detection as well as high computational efficiency. Experimental results on three real SAR image datasets demonstrate the effectiveness of the proposed method.

Yi Wang, Menghan Xia, Lu Qi et al.

Multimodal ambiguity and color bleeding remain challenging in colorization. To tackle these problems, we propose a new GAN-based colorization approach PalGAN, integrated with palette estimation and chromatic attention. To circumvent the multimodality issue, we present a new colorization formulation that estimates a probabilistic palette from the input gray image first, then conducts color assignment conditioned on the palette through a generative model. Further, we handle color bleeding with chromatic attention. It studies color affinities by considering both semantic and intensity correlation. In extensive experiments, PalGAN outperforms state-of-the-arts in quantitative evaluation and visual comparison, delivering notable diverse, contrastive, and edge-preserving appearances. With the palette design, our method enables color transfer between images even with irrelevant contexts.

Tianzi Wang, Shoukang Hu, Jiajun Deng et al.

Automatic recognition of disordered and elderly speech remains highly challenging tasks to date due to data scarcity. Parameter fine-tuning is often used to exploit the large quantities of non-aged and healthy speech pre-trained models, while neural architecture hyper-parameters are set using expert knowledge and remain unchanged. This paper investigates hyper-parameter adaptation for Conformer ASR systems that are pre-trained on the Librispeech corpus before being domain adapted to the DementiaBank elderly and UASpeech dysarthric speech datasets. Experimental results suggest that hyper-parameter adaptation produced word error rate (WER) reductions of 0.45% and 0.67% over parameter-only fine-tuning on DBank and UASpeech tasks respectively. An intuitive correlation is found between the performance improvements by hyper-parameter domain adaptation and the relative utterance length ratio between the source and target domain data.

Jialiang Wang, Haotian Wei, Yi Wang et al.

Human activity recognition (HAR) based on multimodal sensors has become a rapidly growing branch of biometric recognition and artificial intelligence. However, how to fully mine multimodal time series data and effectively learn accurate behavioral features has always been a hot topic in this field. Practical applications also require a well-generalized framework that can quickly process a variety of raw sensor data and learn better feature representations. This paper proposes a universal multi-sensor network (UMSNet) for human activity recognition. In particular, we propose a new lightweight sensor residual block (called LSR block), which improves the performance by reducing the number of activation function and normalization layers, and adding inverted bottleneck structure and grouping convolution. Then, the Transformer is used to extract the relationship of series features to realize the classification and recognition of human activities. Our framework has a clear structure and can be directly applied to various types of multi-modal Time Series Classification (TSC) tasks after simple specialization. Extensive experiments show that the proposed UMSNet outperforms other state-of-the-art methods on two popular multi-sensor human activity recognition datasets (i.e. HHAR dataset and MHEALTH dataset).

Huiping Zhuang, Yuchen Liu, Run He et al.

Online Class Incremental Learning (OCIL) aims to train models incrementally, where data arrive in mini-batches, and previous data are not accessible. A major challenge in OCIL is Catastrophic Forgetting, i.e., the loss of previously learned knowledge. Among existing baselines, replay-based methods show competitive results but requires extra memory for storing exemplars, while exemplar-free (i.e., data need not be stored for replay in production) methods are resource-friendly but often lack accuracy. In this paper, we propose an exemplar-free approach--Forward-only Online Analytic Learning (F-OAL). Unlike traditional methods, F-OAL does not rely on back-propagation and is forward-only, significantly reducing memory usage and computational time. Cooperating with a pre-trained frozen encoder with Feature Fusion, F-OAL only needs to update a linear classifier by recursive least square. This approach simultaneously achieves high accuracy and low resource consumption. Extensive experiments on benchmark datasets demonstrate F-OAL's robust performance in OCIL scenarios. Code is available at https://github.com/liuyuchen-cz/F-OAL.

Haihan Tang, Yi Wang, Lap-Pui Chau

In this paper, we propose a three-stream adaptive fusion network named TAFNet, which uses paired RGB and thermal images for crowd counting. Specifically, TAFNet is divided into one main stream and two auxiliary streams. We combine a pair of RGB and thermal images to constitute the input of main stream. Two auxiliary streams respectively exploit RGB image and thermal image to extract modality-specific features. Besides, we propose an Information Improvement Module (IIM) to fuse the modality-specific features into the main stream adaptively. Experiment results on RGBT-CC dataset show that our method achieves more than 20% improvement on mean average error and root mean squared error compared with state-of-the-art method. The source code will be publicly available at https://github.com/TANGHAIHAN/TAFNet.

Lu Qi, Jason Kuen, Jiuxiang Gu et al.

In instance-level detection tasks (e.g., object detection), reducing input resolution is an easy option to improve runtime efficiency. However, this option traditionally hurts the detection performance much. This paper focuses on boosting the performance of low-resolution models by distilling knowledge from a high- or multi-resolution model. We first identify the challenge of applying knowledge distillation (KD) to teacher and student networks that act on different input resolutions. To tackle it, we explore the idea of spatially aligning feature maps between models of varying input resolutions by shifting feature pyramid positions and introduce aligned multi-scale training to train a multi-scale teacher that can distill its knowledge to a low-resolution student. Further, we propose crossing feature-level fusion to dynamically fuse teacher's multi-resolution features to guide the student better. On several instance-level detection tasks and datasets, the low-resolution models trained via our approach perform competitively with high-resolution models trained via conventional multi-scale training, while outperforming the latter's low-resolution models by 2.1% to 3.6% in terms of mAP. Our code is made publicly available at https://github.com/dvlab-research/MSAD.

Nanyang Yang, Yi Wang, Lap-Pui Chau

The CenterTrack tracking algorithm achieves state-of-the-art tracking performance using a simple detection model and single-frame spatial offsets to localize objects and predict their associations in a single network. However, this joint detection and tracking method still suffers from high identity switches due to the inferior association method. To reduce the high number of identity switches and improve the tracking accuracy, in this paper, we propose to incorporate a simple tracked object bounding box and overlapping prediction based on the current frame onto the CenterTrack algorithm. Specifically, we propose an Intersection over Union (IOU) distance cost matrix in the association step instead of simple point displacement distance. We evaluate our proposed tracker on the MOT17 test dataset, showing that our proposed method can reduce identity switches significantly by 22.6% and obtain a notable improvement of 1.5% in IDF1 compared to the original CenterTrack's under the same tracklet lifetime. The source code is released at https://github.com/Nanyangny/CenterTrack-IOU.

Huiping Zhuang, Yi Wang, Qinglai Liu et al.

Training neural networks with back-propagation (BP) requires a sequential passing of activations and gradients, which forces the network modules to work in a synchronous fashion. This has been recognized as the lockings (i.e., the forward, backward and update lockings) inherited from the BP. In this paper, we propose a fully decoupled training scheme using delayed gradients (FDG) to break all these lockings. The FDG splits a neural network into multiple modules and trains them independently and asynchronously using different workers (e.g., GPUs). We also introduce a gradient shrinking process to reduce the stale gradient effect caused by the delayed gradients. In addition, we prove that the proposed FDG algorithm guarantees a statistical convergence during training. Experiments are conducted by training deep convolutional neural networks to perform classification tasks on benchmark datasets, showing comparable or better results against the state-of-the-art methods as well as the BP in terms of both generalization and acceleration abilities. In particular, we show that the FDG is also able to train very wide networks (e.g., WRN-28-10) and extremely deep networks (e.g., ResNet-1202). Code is available at https://github.com/ZHUANGHP/FDG.

Jiecheng Zhou, Ding Tang, Rong Fu et al.

The burgeoning computational demands for training large language models (LLMs) necessitate efficient methods, including quantized training, which leverages low-bit arithmetic operations to reduce costs. While FP8 precision has shown potential, leveraging FP4 remains challenging due to inherent quantization errors and limited representation capability. Based on the Transformer architecture, we present an FP4 training scheme for LLMs, overcoming these obstacles through mixed-precision quantization strategies tailed for different modules and training stages. This allows us to apply the precision level suitable to distinct components within the model, ensuring that multi-head attention and linear layers are handled appropriately. Our pretraining recipe ensures stability in backpropagation by incorporating fine-grained quantization methods with a target precision training schedule. Experimental results demonstrate that our FP4 training scheme achieves accuracy comparable to BF16 and FP8, with smaller theoretical computational cost. With the advent of next-generation hardware supporting FP4, our method sets the foundation for efficient ultra-low precision training.

Lei Yao, Yi Wang, Yi Zhang et al.

The significance of informative and robust point representations has been widely acknowledged for 3D scene understanding. Despite existing self-supervised pre-training counterparts demonstrating promising performance, the model collapse and structural information deficiency remain prevalent due to insufficient point discrimination difficulty, yielding unreliable expressions and suboptimal performance. In this paper, we present GaussianCross, a novel cross-modal self-supervised 3D representation learning architecture integrating feed-forward 3D Gaussian Splatting (3DGS) techniques to address current challenges. GaussianCross seamlessly converts scale-inconsistent 3D point clouds into a unified cuboid-normalized Gaussian representation without missing details, enabling stable and generalizable pre-training. Subsequently, a tri-attribute adaptive distillation splatting module is incorporated to construct a 3D feature field, facilitating synergetic feature capturing of appearance, geometry, and semantic cues to maintain cross-modal consistency. To validate GaussianCross, we perform extensive evaluations on various benchmarks, including ScanNet, ScanNet200, and S3DIS. In particular, GaussianCross shows a prominent parameter and data efficiency, achieving superior performance through linear probing (<0.1% parameters) and limited data training (1% of scenes) compared to state-of-the-art methods. Furthermore, GaussianCross demonstrates strong generalization capabilities, improving the full fine-tuning accuracy by 9.3% mIoU and 6.1% AP$_{50}$ on ScanNet200 semantic and instance segmentation tasks, respectively, supporting the effectiveness of our approach. The code, weights, and visualizations are publicly available at \href{https://rayyoh.github.io/GaussianCross/}{https://rayyoh.github.io/GaussianCross/}.

Haoran Sun, Chen Cai, Huiping Zhuang et al.

The rapid development of deepfake video technology has not only facilitated artistic creation but also made it easier to spread misinformation. Traditional deepfake video detection (DVD) methods face issues such as a lack of transparency in their principles and insufficient generalization capabilities to cope with evolving forgery techniques. This highlights an urgent need for detectors that can identify forged content and provide verifiable reasoning explanations. This paper proposes the explainable deepfake video detection (EDVD) task and designs the EDVD-LLaMA multimodal, a large language model (MLLM) reasoning framework, which provides traceable reasoning processes alongside accurate detection results and trustworthy explanations. Our approach first incorporates a Spatio-Temporal Subtle Information Tokenization (ST-SIT) to extract and fuse global and local cross-frame deepfake features, providing rich spatio-temporal semantic information input for MLLM reasoning. Second, we construct a Fine-grained Multimodal Chain-of-Thought (Fg-MCoT) mechanism, which introduces facial feature data as hard constraints during the reasoning process to achieve pixel-level spatio-temporal video localization, suppress hallucinated outputs, and enhance the reliability of the chain of thought. In addition, we build an Explainable Reasoning FF++ benchmark dataset (ER-FF++set), leveraging structured data to annotate videos and ensure quality control, thereby supporting dual supervision for reasoning and detection. Extensive experiments demonstrate that EDVD-LLaMA achieves outstanding performance and robustness in terms of detection accuracy, explainability, and its ability to handle cross-forgery methods and cross-dataset scenarios. Compared to previous DVD methods, it provides a more explainable and superior solution. The source code and dataset will be publicly available.

Tianyi Liu, Kejun Wu, Chen Cai et al.

Video signals are vulnerable in multimedia communication and storage systems, as even slight bitstream-domain corruption can lead to significant pixel-domain degradation. To recover faithful spatio-temporal content from corrupted inputs, bitstream-corrupted video recovery has recently emerged as a challenging and understudied task. However, existing methods require time-consuming and labor-intensive annotation of corrupted regions for each corrupted video frame, resulting in a large workload in practice. In addition, high-quality recovery remains difficult as part of the local residual information in corrupted frames may mislead feature completion and successive content recovery. In this paper, we propose the first blind bitstream-corrupted video recovery framework that integrates visual foundation models with a recovery model, which is adapted to different types of corruption and bitstream-level prompts. Within the framework, the proposed Detect Any Corruption (DAC) model leverages the rich priors of the visual foundation model while incorporating bitstream and corruption knowledge to enhance corruption localization and blind recovery. Additionally, we introduce a novel Corruption-aware Feature Completion (CFC) module, which adaptively processes residual contributions based on high-level corruption understanding. With VFM-guided hierarchical feature augmentation and high-level coordination in a mixture-of-residual-experts (MoRE) structure, our method suppresses artifacts and enhances informative residuals. Comprehensive evaluations show that the proposed method achieves outstanding performance in bitstream-corrupted video recovery without requiring a manually labeled mask sequence. The demonstrated effectiveness will help to realize improved user experience, wider application scenarios, and more reliable multimedia communication and storage systems.

Yichen Zhu, Yuqin Zhu, Jie Du et al.

The vision transformer splits each image into a sequence of tokens with fixed length and processes the tokens in the same way as words in natural language processing. More tokens normally lead to better performance but considerably increased computational cost. Motivated by the proverb "A picture is worth a thousand words" we aim to accelerate the ViT model by making a long image short. To this end, we propose a novel approach to assign token length adaptively during inference. Specifically, we first train a ViT model, called Resizable-ViT (ReViT), that can process any given input with diverse token lengths. Then, we retrieve the "token-length label" from ReViT and use it to train a lightweight Token-Length Assigner (TLA). The token-length labels are the smallest number of tokens to split an image that the ReViT can make the correct prediction, and TLA is learned to allocate the optimal token length based on these labels. The TLA enables the ReViT to process the image with the minimum sufficient number of tokens during inference. Thus, the inference speed is boosted by reducing the token numbers in the ViT model. Our approach is general and compatible with modern vision transformer architectures and can significantly reduce computational expanse. We verified the effectiveness of our methods on multiple representative ViT models (DeiT, LV-ViT, and TimesFormer) across two tasks (image classification and action recognition).

Yi Wang, Yuchen He, Xutian Deng et al.

With the potential applications of capsule robots in medical endoscopy, accurate dynamic control of the capsule robot is becoming more and more important. In the scale of a capsule robot, the friction between capsule and the environment plays an essential role in the dynamic model, which is usually difficult to model beforehand. In the paper, a tethered capsule robot system driven by a robot manipulator is built, where a strong magnetic Halbach array is mounted on the robot's end-effector to adjust the state of the capsule. To increase the control accuracy, the friction between capsule and the environment is learned with demonstrated trajectories. With the learned friction model, experimental results demonstrate an improvement of 5.6% in terms of tracking error.

Lisha Tang, Yi Wang, Lap-Pui Chau

Vehicle re-identification (Re-ID) aims to retrieve images with the same vehicle ID across different cameras. Current part-level feature learning methods typically detect vehicle parts via uniform division, outside tools, or attention modeling. However, such part features often require expensive additional annotations and cause sub-optimal performance in case of unreliable part mask predictions. In this paper, we propose a weakly-supervised Part-Attention Network (PANet) and Part-Mentored Network (PMNet) for Vehicle Re-ID. Firstly, PANet localizes vehicle parts via part-relevant channel recalibration and cluster-based mask generation without vehicle part supervisory information. Secondly, PMNet leverages teacher-student guided learning to distill vehicle part-specific features from PANet and performs multi-scale global-part feature extraction. During inference, PMNet can adaptively extract discriminative part features without part localization by PANet, preventing unstable part mask predictions. We address this Re-ID issue as a multi-task problem and adopt Homoscedastic Uncertainty to learn the optimal weighing of ID losses. Experiments are conducted on two public benchmarks, showing that our approach outperforms recent methods, which require no extra annotations by an average increase of 3.0% in CMC@5 on VehicleID and over 1.4% in mAP on VeRi776. Moreover, our method can extend to the occluded vehicle Re-ID task and exhibits good generalization ability.

Yunhao Zhou, Yi Wang, Lap-Pui Chau

Re-ranking utilizes contextual information to optimize the initial ranking list of person or vehicle re-identification (re-ID), which boosts the retrieval performance at post-processing steps. This paper proposes a re-ranking network to predict the correlations between the probe and top-ranked neighbor samples. Specifically, all the feature embeddings of query and gallery images are expanded and enhanced by a linear combination of their neighbors, with the correlation prediction serving as discriminative combination weights. The combination process is equivalent to moving independent embeddings toward the identity centers, improving cluster compactness. For correlation prediction, we first aggregate the contextual information for probe's k-nearest neighbors via the Transformer encoder. Then, we distill and refine the probe-related features into the Contextual Memory cell via attention mechanism. Like humans that retrieve images by not only considering probe images but also memorizing the retrieved ones, the Contextual Memory produces multi-view descriptions for each instance. Finally, the neighbors are reconstructed with features fetched from the Contextual Memory, and a binary classifier predicts their correlations with the probe. Experiments on six widely-used person and vehicle re-ID benchmarks demonstrate the effectiveness of the proposed method. Especially, our method surpasses the state-of-the-art re-ranking approaches on large-scale datasets by a significant margin, i.e., with an average 4.83% CMC@1 and 14.83% mAP improvements on VERI-Wild, MSMT17, and VehicleID datasets.

Arman Maesumi, Mingkang Zhu, Yi Wang et al.

This paper presents a novel patch-based adversarial attack pipeline that trains adversarial patches on 3D human meshes. We sample triangular faces on a reference human mesh, and create an adversarial texture atlas over those faces. The adversarial texture is transferred to human meshes in various poses, which are rendered onto a collection of real-world background images. Contrary to the traditional patch-based adversarial attacks, where prior work attempts to fool trained object detectors using appended adversarial patches, this new form of attack is mapped into the 3D object world and back-propagated to the texture atlas through differentiable rendering. As such, the adversarial patch is trained under deformation consistent with real-world materials. In addition, and unlike existing adversarial patches, our new 3D adversarial patch is shown to fool state-of-the-art deep object detectors robustly under varying views, potentially leading to an attacking scheme that is persistently strong in the physical world.

Menghan Xia, Yi Wang, Chu Han et al.

As a generic modeling tool, Convolutional Neural Networks (CNNs) have been widely employed in image generation and translation tasks. However, when fed with a flat input, current CNN models may fail to generate vivid results due to the spatially shared convolution kernels. We call it the flatness degradation of CNNs. Unfortunately, such degradation is the greatest obstacles to generate a spatially-variant output from a flat input, which has been barely discussed in the previous literature. To tackle this problem, we propose a model agnostic solution, i.e. Noise Incentive Block (NIB), which serves as a generic plug-in for any CNN generation model. The key idea is to break the flat input condition while keeping the intactness of the original information. Specifically, the NIB perturbs the input data symmetrically with a noise map and reassembles them in the feature domain as driven by the objective function. Extensive experiments show that existing CNN models equipped with NIB survive from the flatness degradation and are able to generate visually better results with richer details in some specific image generation tasks given flat inputs, e.g. semantic image synthesis, data-hidden image generation, and deep neural dithering.

Yi Wang, Zhen-Peng Bian, Yunhao Zhou et al.

In this paper, we propose a real-time and accurate automatic license plate recognition (ALPR) approach. Our study illustrates the outstanding design of ALPR with four insights: (1) the resampling-based cascaded framework is beneficial to both speed and accuracy; (2) the highly efficient license plate recognition should abundant additional character segmentation and recurrent neural network (RNN), but adopt a plain convolutional neural network (CNN); (3) in the case of CNN, taking advantage of vertex information on license plates improves the recognition performance; and (4) the weight-sharing character classifier addresses the lack of training images in small-scale datasets. Based on these insights, we propose a novel ALPR approach, termed VSNet. Specifically, VSNet includes two CNNs, i.e., VertexNet for license plate detection and SCR-Net for license plate recognition, integrated in a resampling-based cascaded manner. In VertexNet, we propose an efficient integration block to extract the spatial features of license plates. With vertex supervisory information, we propose a vertex-estimation branch in VertexNet such that license plates can be rectified as the input images of SCR-Net. In SCR-Net, we introduce a horizontal encoding technique for left-to-right feature extraction and propose a weight-sharing classifier for character recognition. Experimental results show that the proposed VSNet outperforms state-of-the-art methods by more than 50% relative improvement on error rate, achieving > 99% recognition accuracy on CCPD and AOLP datasets with 149 FPS inference speed. Moreover, our method illustrates an outstanding generalization capability when evaluated on the unseen PKUData and CLPD datasets.

Hang Zhang, Jinwei Zhang, Rongguang Wang et al.

Recently, 3D medical image reconstruction (MIR) and segmentation (MIS) based on deep neural networks have been developed with promising results, and attention mechanism has been further designed to capture global contextual information for performance enhancement. However, the large size of 3D volume images poses a great computational challenge to traditional attention methods. In this paper, we propose a folded attention (FA) approach to improve the computational efficiency of traditional attention methods on 3D medical images. The main idea is that we apply tensor folding and unfolding operations with four permutations to build four small sub-affinity matrices to approximate the original affinity matrix. Through four consecutive sub-attention modules of FA, each element in the feature tensor can aggregate spatial-channel information from all other elements. Compared to traditional attention methods, with moderate improvement of accuracy, FA can substantially reduce the computational complexity and GPU memory consumption. We demonstrate the superiority of our method on two challenging tasks for 3D MIR and MIS, which are quantitative susceptibility mapping and multiple sclerosis lesion segmentation.

Zhiyu Zhu, Zhen-Peng Bian, Junhui Hou et al.

Various architectures (such as GoogLeNets, ResNets, and DenseNets) have been proposed. However, the existing networks usually suffer from either redundancy of convolutional layers or insufficient utilization of parameters. To handle these challenging issues, we propose Micro-Dense Nets, a novel architecture with global residual learning and local micro-dense aggregations. Specifically, residual learning aims to efficiently retrieve features from different convolutional blocks, while the micro-dense aggregation is able to enhance each block and avoid redundancy of convolutional layers by lessening residual aggregations. Moreover, the proposed micro-dense architecture has two characteristics: pyramidal multi-level feature learning which can widen the deeper layer in a block progressively, and dimension cardinality adaptive convolution which can balance each layer using linearly increasing dimension cardinality. The experimental results over three datasets (i.e., CIFAR-10, CIFAR-100, and ImageNet-1K) demonstrate that the proposed Micro-Dense Net with only 4M parameters can achieve higher classification accuracy than state-of-the-art networks, while being 12.1$\times$ smaller depends on the number of parameters. In addition, our micro-dense block can be integrated with neural architecture search based models to boost their performance, validating the advantage of our architecture. We believe our design and findings will be beneficial to the DNN community.

Yi Wang, Ying-Cong Chen, Xiangyu Zhang et al.

Traditional convolution-based generative adversarial networks synthesize images based on hierarchical local operations, where long-range dependency relation is implicitly modeled with a Markov chain. It is still not sufficient for categories with complicated structures. In this paper, we characterize long-range dependence with attentive normalization (AN), which is an extension to traditional instance normalization. Specifically, the input feature map is softly divided into several regions based on its internal semantic similarity, which are respectively normalized. It enhances consistency between distant regions with semantic correspondence. Compared with self-attention GAN, our attentive normalization does not need to measure the correlation of all locations, and thus can be directly applied to large-size feature maps without much computational burden. Extensive experiments on class-conditional image generation and semantic inpainting verify the efficacy of our proposed module.

Yi Wang, Zhen-Peng Bian, Junhui Hou et al.

Regularization is commonly used for alleviating overfitting in machine learning. For convolutional neural networks (CNNs), regularization methods, such as DropBlock and Shake-Shake, have illustrated the improvement in the generalization performance. However, these methods lack a self-adaptive ability throughout training. That is, the regularization strength is fixed to a predefined schedule, and manual adjustments are required to adapt to various network architectures. In this paper, we propose a dynamic regularization method for CNNs. Specifically, we model the regularization strength as a function of the training loss. According to the change of the training loss, our method can dynamically adjust the regularization strength in the training procedure, thereby balancing the underfitting and overfitting of CNNs. With dynamic regularization, a large-scale model is automatically regularized by the strong perturbation, and vice versa. Experimental results show that the proposed method can improve the generalization capability on off-the-shelf network architectures and outperform state-of-the-art regularization methods.

Xin Tao, Hongyun Gao, Yi Wang et al.

In single image deblurring, the "coarse-to-fine" scheme, i.e. gradually restoring the sharp image on different resolutions in a pyramid, is very successful in both traditional optimization-based methods and recent neural-network-based approaches. In this paper, we investigate this strategy and propose a Scale-recurrent Network (SRN-DeblurNet) for this deblurring task. Compared with the many recent learning-based approaches in [25], it has a simpler network structure, a smaller number of parameters and is easier to train. We evaluate our method on large-scale deblurring datasets with complex motion. Results show that our method can produce better quality results than state-of-the-arts, both quantitatively and qualitatively.