Jiawei Zhang

Ling-Hao Chen, Jiawei Zhang, Yewen Li et al. · tsinghua

Human motion prediction is a classical problem in computer vision and computer graphics, which has a wide range of practical applications. Previous effects achieve great empirical performance based on an encoding-decoding style. The methods of this style work by first encoding previous motions to latent representations and then decoding the latent representations into predicted motions. However, in practice, they are still unsatisfactory due to several issues, including complicated loss constraints, cumbersome training processes, and scarce switch of different categories of motions in prediction. In this paper, to address the above issues, we jump out of the foregoing style and propose a novel framework from a new perspective. Specifically, our framework works in a masked completion fashion. In the training stage, we learn a motion diffusion model that generates motions from random noise. In the inference stage, with a denoising procedure, we make motion prediction conditioning on observed motions to output more continuous and controllable predictions. The proposed framework enjoys promising algorithmic properties, which only needs one loss in optimization and is trained in an end-to-end manner. Additionally, it accomplishes the switch of different categories of motions effectively, which is significant in realistic tasks, e.g., the animation task. Comprehensive experiments on benchmarks confirm the superiority of the proposed framework. The project page is available at https://lhchen.top/Human-MAC.

Jiawei Zhang, Xiang Wang, Xiao Bai et al.

Despite recent stereo matching networks achieving impressive performance given sufficient training data, they suffer from domain shifts and generalize poorly to unseen domains. We argue that maintaining feature consistency between matching pixels is a vital factor for promoting the generalization capability of stereo matching networks, which has not been adequately considered. Here we address this issue by proposing a simple pixel-wise contrastive learning across the viewpoints. The stereo contrastive feature loss function explicitly constrains the consistency between learned features of matching pixel pairs which are observations of the same 3D points. A stereo selective whitening loss is further introduced to better preserve the stereo feature consistency across domains, which decorrelates stereo features from stereo viewpoint-specific style information. Counter-intuitively, the generalization of feature consistency between two viewpoints in the same scene translates to the generalization of stereo matching performance to unseen domains. Our method is generic in nature as it can be easily embedded into existing stereo networks and does not require access to the samples in the target domain. When trained on synthetic data and generalized to four real-world testing sets, our method achieves superior performance over several state-of-the-art networks.

Jiahe Li, Jiawei Zhang, Xiao Bai et al.

This paper presents ER-NeRF, a novel conditional Neural Radiance Fields (NeRF) based architecture for talking portrait synthesis that can concurrently achieve fast convergence, real-time rendering, and state-of-the-art performance with small model size. Our idea is to explicitly exploit the unequal contribution of spatial regions to guide talking portrait modeling. Specifically, to improve the accuracy of dynamic head reconstruction, a compact and expressive NeRF-based Tri-Plane Hash Representation is introduced by pruning empty spatial regions with three planar hash encoders. For speech audio, we propose a Region Attention Module to generate region-aware condition feature via an attention mechanism. Different from existing methods that utilize an MLP-based encoder to learn the cross-modal relation implicitly, the attention mechanism builds an explicit connection between audio features and spatial regions to capture the priors of local motions. Moreover, a direct and fast Adaptive Pose Encoding is introduced to optimize the head-torso separation problem by mapping the complex transformation of the head pose into spatial coordinates. Extensive experiments demonstrate that our method renders better high-fidelity and audio-lips synchronized talking portrait videos, with realistic details and high efficiency compared to previous methods.

Zhouxia Wang, Jiawei Zhang, Tianshui Chen et al.

Blind face restoration aims at recovering high-quality face images from those with unknown degradations. Current algorithms mainly introduce priors to complement high-quality details and achieve impressive progress. However, most of these algorithms ignore abundant contextual information in the face and its interplay with the priors, leading to sub-optimal performance. Moreover, they pay less attention to the gap between the synthetic and real-world scenarios, limiting the robustness and generalization to real-world applications. In this work, we propose RestoreFormer++, which on the one hand introduces fully-spatial attention mechanisms to model the contextual information and the interplay with the priors, and on the other hand, explores an extending degrading model to help generate more realistic degraded face images to alleviate the synthetic-to-real-world gap. Compared with current algorithms, RestoreFormer++ has several crucial benefits. First, instead of using a multi-head self-attention mechanism like the traditional visual transformer, we introduce multi-head cross-attention over multi-scale features to fully explore spatial interactions between corrupted information and high-quality priors. In this way, it can facilitate RestoreFormer++ to restore face images with higher realness and fidelity. Second, in contrast to the recognition-oriented dictionary, we learn a reconstruction-oriented dictionary as priors, which contains more diverse high-quality facial details and better accords with the restoration target. Third, we introduce an extending degrading model that contains more realistic degraded scenarios for training data synthesizing, and thus helps to enhance the robustness and generalization of our RestoreFormer++ model. Extensive experiments show that RestoreFormer++ outperforms state-of-the-art algorithms on both synthetic and real-world datasets.

Aiqing Zhu, Beibei Zhu, Jiawei Zhang et al.

We propose volume-preserving networks (VPNets) for learning unknown source-free dynamical systems using trajectory data. We propose three modules and combine them to obtain two network architectures, coined R-VPNet and LA-VPNet. The distinct feature of the proposed models is that they are intrinsic volume-preserving. In addition, the corresponding approximation theorems are proved, which theoretically guarantee the expressivity of the proposed VPNets to learn source-free dynamics. The effectiveness, generalization ability and structure-preserving property of the VP-Nets are demonstrated by numerical experiments.

Hui Hu, Jiawei Zhang, Zhen Han

Heterogeneous face re-identification, namely matching heterogeneous faces across disjoint visible light (VIS) and near-infrared (NIR) cameras, has become an important problem in video surveillance application. However, the large domain discrepancy between heterogeneous NIR-VIS faces makes the performance of face re-identification degraded dramatically. To solve this problem, a multimodal fusion ranking optimization algorithm for heterogeneous face re-identification is proposed in this paper. Firstly, we design a heterogeneous face translation network to obtain multimodal face pairs, including NIR-VIS/NIR-NIR/VIS-VIS face pairs, through mutual transformation between NIR-VIS faces. Secondly, we propose linear and non-linear fusion strategies to aggregate initial ranking lists of multimodal face pairs and acquire the optimized re-ranked list based on modal complementarity. The experimental results show that the proposed multimodal fusion ranking optimization algorithm can effectively utilize the complementarity and outperforms some relative methods on the SCface dataset.

Yiming Huang, Aozhe Jia, Xiaodan Zhang et al.

Attention-based transformer models have achieved remarkable progress in multi-modal tasks, such as visual question answering. The explainability of attention-based methods has recently attracted wide interest as it can explain the inner changes of attention tokens by accumulating relevancy across attention layers. Current methods simply update relevancy by equally accumulating the token relevancy before and after the attention processes. However, the importance of token values is usually different during relevance accumulation. In this paper, we propose a weighted relevancy strategy, which takes the importance of token values into consideration, to reduce distortion when equally accumulating relevance. To evaluate our method, we propose a unified CLIP-based two-stage model, named CLIPmapper, to process Vision-and-Language tasks through CLIP encoder and a following mapper. CLIPmapper consists of self-attention, cross-attention, single-modality, and cross-modality attention, thus it is more suitable for evaluating our generic explainability method. Extensive perturbation tests on visual question answering and image captioning validate that our explainability method outperforms existing methods.

Xinhao Xiang, Simon Dräger, Jiawei Zhang

Good 3D object detection performance from LiDAR-Camera sensors demands seamless feature alignment and fusion strategies. We propose the 3DifFusionDet framework in this paper, which structures 3D object detection as a denoising diffusion process from noisy 3D boxes to target boxes. In this framework, ground truth boxes diffuse in a random distribution for training, and the model learns to reverse the noising process. During inference, the model gradually refines a set of boxes that were generated at random to the outcomes. Under the feature align strategy, the progressive refinement method could make a significant contribution to robust LiDAR-Camera fusion. The iterative refinement process could also demonstrate great adaptability by applying the framework to various detecting circumstances where varying levels of accuracy and speed are required. Extensive experiments on KITTI, a benchmark for real-world traffic object identification, revealed that 3DifFusionDet is able to perform favorably in comparison to earlier, well-respected detectors.

Jiahe Li, Jiawei Zhang, Xiao Bai et al.

Radiance fields have demonstrated impressive performance in synthesizing novel views from sparse input views, yet prevailing methods suffer from high training costs and slow inference speed. This paper introduces DNGaussian, a depth-regularized framework based on 3D Gaussian radiance fields, offering real-time and high-quality few-shot novel view synthesis at low costs. Our motivation stems from the highly efficient representation and surprising quality of the recent 3D Gaussian Splatting, despite it will encounter a geometry degradation when input views decrease. In the Gaussian radiance fields, we find this degradation in scene geometry primarily lined to the positioning of Gaussian primitives and can be mitigated by depth constraint. Consequently, we propose a Hard and Soft Depth Regularization to restore accurate scene geometry under coarse monocular depth supervision while maintaining a fine-grained color appearance. To further refine detailed geometry reshaping, we introduce Global-Local Depth Normalization, enhancing the focus on small local depth changes. Extensive experiments on LLFF, DTU, and Blender datasets demonstrate that DNGaussian outperforms state-of-the-art methods, achieving comparable or better results with significantly reduced memory cost, a $25 \times$ reduction in training time, and over $3000 \times$ faster rendering speed.

Xinhao Xiang, Jiawei Zhang

For 3D object detection, both camera and lidar have been demonstrated to be useful sensory devices for providing complementary information about the same scenery with data representations in different modalities, e.g., 2D RGB image vs 3D point cloud. An effective representation learning and fusion of such multi-modal sensor data is necessary and critical for better 3D object detection performance. To solve the problem, in this paper, we will introduce a novel vision transformer-based 3D object detection model, namely FusionViT. Different from the existing 3D object detection approaches, FusionViT is a pure-ViT based framework, which adopts a hierarchical architecture by extending the transformer model to embed both images and point clouds for effective representation learning. Such multi-modal data embedding representations will be further fused together via a fusion vision transformer model prior to feeding the learned features to the object detection head for both detection and localization of the 3D objects in the input scenery. To demonstrate the effectiveness of FusionViT, extensive experiments have been done on real-world traffic object detection benchmark datasets KITTI and Waymo Open. Notably, our FusionViT model can achieve state-of-the-art performance and outperforms not only the existing baseline methods that merely rely on camera images or lidar point clouds, but also the latest multi-modal image-point cloud deep fusion approaches.

Zhilu Zhang, Haolin Wang, Ming Liu et al.

Learning RAW-to-sRGB mapping has drawn increasing attention in recent years, wherein an input raw image is trained to imitate the target sRGB image captured by another camera. However, the severe color inconsistency makes it very challenging to generate well-aligned training pairs of input raw and target sRGB images. While learning with inaccurately aligned supervision is prone to causing pixel shift and producing blurry results. In this paper, we circumvent such issue by presenting a joint learning model for image alignment and RAW-to-sRGB mapping. To diminish the effect of color inconsistency in image alignment, we introduce to use a global color mapping (GCM) module to generate an initial sRGB image given the input raw image, which can keep the spatial location of the pixels unchanged, and the target sRGB image is utilized to guide GCM for converting the color towards it. Then a pre-trained optical flow estimation network (e.g., PWC-Net) is deployed to warp the target sRGB image to align with the GCM output. To alleviate the effect of inaccurately aligned supervision, the warped target sRGB image is leveraged to learn RAW-to-sRGB mapping. When training is done, the GCM module and optical flow network can be detached, thereby bringing no extra computation cost for inference. Experiments show that our method performs favorably against state-of-the-arts on ZRR and SR-RAW datasets. With our joint learning model, a light-weight backbone can achieve better quantitative and qualitative performance on ZRR dataset. Codes are available at https://github.com/cszhilu1998/RAW-to-sRGB.

Jiawei Zhang, Jiahe Li, Lei Huang et al.

With advancements in domain generalized stereo matching networks, models pre-trained on synthetic data demonstrate strong robustness to unseen domains. However, few studies have investigated the robustness after fine-tuning them in real-world scenarios, during which the domain generalization ability can be seriously degraded. In this paper, we explore fine-tuning stereo matching networks without compromising their robustness to unseen domains. Our motivation stems from comparing Ground Truth (GT) versus Pseudo Label (PL) for fine-tuning: GT degrades, but PL preserves the domain generalization ability. Empirically, we find the difference between GT and PL implies valuable information that can regularize networks during fine-tuning. We also propose a framework to utilize this difference for fine-tuning, consisting of a frozen Teacher, an exponential moving average (EMA) Teacher, and a Student network. The core idea is to utilize the EMA Teacher to measure what the Student has learned and dynamically improve GT and PL for fine-tuning. We integrate our framework with state-of-the-art networks and evaluate its effectiveness on several real-world datasets. Extensive experiments show that our method effectively preserves the domain generalization ability during fine-tuning.

Yingying Wang, Yun Xiong, Xixi Wu et al.

Drug combinations can cause adverse drug-drug interactions(DDIs). Identifying specific effects is crucial for developing safer therapies. Previous works on DDI event prediction have typically been limited to using labels of specific events as supervision, which renders them insufficient to address two significant challenges: (1) the bias caused by \textbf{highly imbalanced event distribution} where certain interaction types are vastly under-represented. (2) the \textbf{scarcity of labeled data for rare events}, a pervasive issue where rare yet potentially critical interactions are often overlooked or under-explored due to limited available data. In response, we offer ``DDIPrompt'', an innovative solution inspired by the recent advancements in graph prompt learning. Our framework aims to address these issues by leveraging the intrinsic knowledge from pre-trained models, which can be efficiently deployed with minimal downstream data. Specifically, to solve the first challenge, DDIPrompt features a hierarchical pre-training strategy to foster a generalized and comprehensive understanding of drug properties. It captures intra-molecular structures through augmented links based on structural proximity between drugs, further learns inter-molecular interactions emphasizing edge connections rather than concrete catagories. For the second challenge, we implement a prototype-enhanced prompting mechanism during inference. This mechanism, refined by few-shot examples from each category, effectively harnesses the rich pre-training knowledge to enhance prediction accuracy, particularly for these rare but crucial interactions. Extensive experiments on two benchmark datasets demonstrate DDIPrompt's SOTA performance, especially for those rare DDI events.

Lin Yuan, Xiaowan Li, Yan Zhang et al.

Advancements in image generation technologies have raised significant concerns about their potential misuse, such as producing misinformation and deepfakes. Therefore, there is an urgent need for effective methods to detect AI-generated images (AIGI). Despite progress in AIGI detection, achieving reliable performance across diverse generation models and scenes remains challenging due to the lack of source-invariant features and limited generalization capabilities in existing methods. In this work, we explore the potential of using image entropy as a cue for AIGI detection and propose Multi-granularity Local Entropy Patterns (MLEP), a set of entropy feature maps computed across shuffled small patches over multiple image scaled. MLEP comprehensively captures pixel relationships across dimensions and scales while significantly disrupting image semantics, reducing potential content bias. Leveraging MLEP, a robust CNN-based classifier for AIGI detection can be trained. Extensive experiments conducted in an open-world scenario, evaluating images synthesized by 32 distinct generative models, demonstrate significant improvements over state-of-the-art methods in both accuracy and generalization.

Xiaowe Xu, Jiawei Zhang, Jinglan Liu et al.

As one of the most commonly ordered imaging tests, computed tomography (CT) scan comes with inevitable radiation exposure that increases the cancer risk to patients. However, CT image quality is directly related to radiation dose, thus it is desirable to obtain high-quality CT images with as little dose as possible. CT image denoising tries to obtain high dose like high-quality CT images (domain X) from low dose low-quality CTimages (domain Y), which can be treated as an image-to-image translation task where the goal is to learn the transform between a source domain X (noisy images) and a target domain Y (clean images). In this paper, we propose a multi-cycle-consistent adversarial network (MCCAN) that builds intermediate domains and enforces both local and global cycle-consistency for edge denoising of CT images. The global cycle-consistency couples all generators together to model the whole denoising process, while the local cycle-consistency imposes effective supervision on the process between adjacent domains. Experiments show that both local and global cycle-consistency are important for the success of MCCAN, which outperformsCCADN in terms of denoising quality with slightly less computation resource consumption.

Jiawei Zhang, Yanchun Zhang, Xiaowei Xu

Semantic image segmentation aims to obtain object labels with precise boundaries, which usually suffers from overfitting. Recently, various data augmentation strategies like regional dropout and mix strategies have been proposed to address the problem. These strategies have proved to be effective for guiding the model to attend on less discriminative parts. However, current strategies operate at the image level, and objects and the background are coupled. Thus, the boundaries are not well augmented due to the fixed semantic scenario. In this paper, we propose ObjectAug to perform object-level augmentation for semantic image segmentation. ObjectAug first decouples the image into individual objects and the background using the semantic labels. Next, each object is augmented individually with commonly used augmentation methods (e.g., scaling, shifting, and rotation). Then, the black area brought by object augmentation is further restored using image inpainting. Finally, the augmented objects and background are assembled as an augmented image. In this way, the boundaries can be fully explored in the various semantic scenarios. In addition, ObjectAug can support category-aware augmentation that gives various possibilities to objects in each category, and can be easily combined with existing image-level augmentation methods to further boost performance. Comprehensive experiments are conducted on both natural image and medical image datasets. Experiment results demonstrate that our ObjectAug can evidently improve segmentation performance.

Jiawei Zhang, Yang Wang, Piero Molino et al.

Interpretation and diagnosis of machine learning models have gained renewed interest in recent years with breakthroughs in new approaches. We present Manifold, a framework that utilizes visual analysis techniques to support interpretation, debugging, and comparison of machine learning models in a more transparent and interactive manner. Conventional techniques usually focus on visualizing the internal logic of a specific model type (i.e., deep neural networks), lacking the ability to extend to a more complex scenario where different model types are integrated. To this end, Manifold is designed as a generic framework that does not rely on or access the internal logic of the model and solely observes the input (i.e., instances or features) and the output (i.e., the predicted result and probability distribution). We describe the workflow of Manifold as an iterative process consisting of three major phases that are commonly involved in the model development and diagnosis process: inspection (hypothesis), explanation (reasoning), and refinement (verification). The visual components supporting these tasks include a scatterplot-based visual summary that overviews the models' outcome and a customizable tabular view that reveals feature discrimination. We demonstrate current applications of the framework on the classification and regression tasks and discuss other potential machine learning use scenarios where Manifold can be applied.

Jiawei Zhang, Limeng Cui, Fisher B. Gouza

In this paper, we propose to provide a general ensemble learning framework based on deep learning models. Given a group of unit models, the proposed deep ensemble learning framework will effectively combine their learning results via a multilayered ensemble model. In the case when the unit model mathematical mappings are bounded, sigmoidal and discriminatory, we demonstrate that the deep ensemble learning framework can achieve a universal approximation of any functions from the input space to the output space. Meanwhile, to achieve such a performance, the deep ensemble learning framework also impose a strict constraint on the number of involved unit models. According to the theoretic proof provided in this paper, given the input feature space of dimension d, the required unit model number will be 2d, if the ensemble model involves one single layer. Furthermore, as the ensemble component goes deeper, the number of required unit model is proved to be lowered down exponentially.

Yicun Liu, Jimmy Ren, Jianbo Liu et al.

Artistic style transfer can be thought as a process to generate different versions of abstraction of the original image. However, most of the artistic style transfer operators are not optimized for human faces thus mainly suffers from two undesirable features when applying them to selfies. First, the edges of human faces may unpleasantly deviate from the ones in the original image. Second, the skin color is far from faithful to the original one which is usually problematic in producing quality selfies. In this paper, we take a different approach and formulate this abstraction process as a gradient domain learning problem. We aim to learn a type of abstraction which not only achieves the specified artistic style but also circumvents the two aforementioned drawbacks thus highly applicable to selfie photography. We also show that our method can be directly generalized to videos with high inter-frame consistency. Our method is also robust to non-selfie images, and the generalization to various kinds of real-life scenes is discussed. We will make our code publicly available.

Wenqi Ren, Lin Ma, Jiawei Zhang et al.

In this paper, we propose an efficient algorithm to directly restore a clear image from a hazy input. The proposed algorithm hinges on an end-to-end trainable neural network that consists of an encoder and a decoder. The encoder is exploited to capture the context of the derived input images, while the decoder is employed to estimate the contribution of each input to the final dehazed result using the learned representations attributed to the encoder. The constructed network adopts a novel fusion-based strategy which derives three inputs from an original hazy image by applying White Balance (WB), Contrast Enhancing (CE), and Gamma Correction (GC). We compute pixel-wise confidence maps based on the appearance differences between these different inputs to blend the information of the derived inputs and preserve the regions with pleasant visibility. The final dehazed image is yielded by gating the important features of the derived inputs. To train the network, we introduce a multi-scale approach such that the halo artifacts can be avoided. Extensive experimental results on both synthetic and real-world images demonstrate that the proposed algorithm performs favorably against the state-of-the-art algorithms.

Jiawei Zhang, Jianbo Jiao, Mingliang Chen et al.

3D hand pose tracking/estimation will be very important in the next generation of human-computer interaction. Most of the currently available algorithms rely on low-cost active depth sensors. However, these sensors can be easily interfered by other active sources and require relatively high power consumption. As a result, they are currently not suitable for outdoor environments and mobile devices. This paper aims at tracking/estimating hand poses using passive stereo which avoids these limitations. A benchmark with 18,000 stereo image pairs and 18,000 depth images captured from different scenarios and the ground-truth 3D positions of palm and finger joints (obtained from the manual label) is thus proposed. This paper demonstrates that the performance of the state-of-the art tracking/estimation algorithms can be maintained with most stereo matching algorithms on the proposed benchmark, as long as the hand segmentation is correct. As a result, a novel stereo-based hand segmentation algorithm specially designed for hand tracking/estimation is proposed. The quantitative evaluation demonstrates that the proposed algorithm is suitable for the state-of-the-art hand pose tracking/estimation algorithms and the tracking quality is comparable to the use of active depth sensors under different challenging scenarios.