Zhuo Su

Zhuo Su, Max Welling, Matti Pietikäinen et al.

Efficiency and robustness are increasingly needed for applications on 3D point clouds, with the ubiquitous use of edge devices in scenarios like autonomous driving and robotics, which often demand real-time and reliable responses. The paper tackles the challenge by designing a general framework to construct 3D learning architectures with SO(3) equivariance and network binarization. However, a naive combination of equivariant networks and binarization either causes sub-optimal computational efficiency or geometric ambiguity. We propose to locate both scalar and vector features in our networks to avoid both cases. Precisely, the presence of scalar features makes the major part of the network binarizable, while vector features serve to retain rich structural information and ensure SO(3) equivariance. The proposed approach can be applied to general backbones like PointNet and DGCNN. Meanwhile, experiments on ModelNet40, ShapeNet, and the real-world dataset ScanObjectNN, demonstrated that the method achieves a great trade-off between efficiency, rotation robustness, and accuracy. The codes are available at https://github.com/zhuoinoulu/svnet.

Xin Chen, Zhuo Su, Lingbo Yang et al. · tencent-ai

Motion capture from a monocular video is fundamental and crucial for us humans to naturally experience and interact with each other in Virtual Reality (VR) and Augmented Reality (AR). However, existing methods still struggle with challenging cases involving self-occlusion and complex poses due to the lack of effective motion prior modeling. In this paper, we present a novel variational motion prior (VMP) learning approach for video-based motion capture to resolve the above issue. Instead of directly building the correspondence between the video and motion domain, We propose to learn a generic latent space for capturing the prior distribution of all natural motions, which serve as the basis for subsequent video-based motion capture tasks. To improve the generalization capacity of prior space, we propose a transformer-based variational autoencoder pretrained over marker-based 3D mocap data, with a novel style-mapping block to boost the generation quality. Afterward, a separate video encoder is attached to the pretrained motion generator for end-to-end fine-tuning over task-specific video datasets. Compared to existing motion prior models, our VMP model serves as a motion rectifier that can effectively reduce temporal jittering and failure modes in frame-wise pose estimation, leading to temporally stable and visually realistic motion capture results. Furthermore, our VMP-based framework models motion at sequence level and can directly generate motion clips in the forward pass, achieving real-time motion capture during inference. Extensive experiments over both public datasets and in-the-wild videos have demonstrated the efficacy and generalization capability of our framework.

Zhuo Su, Jiehua Zhang, Tianpeng Liu et al.

This paper proposes a novel module called middle spectrum grouped convolution (MSGC) for efficient deep convolutional neural networks (DCNNs) with the mechanism of grouped convolution. It explores the broad "middle spectrum" area between channel pruning and conventional grouped convolution. Compared with channel pruning, MSGC can retain most of the information from the input feature maps due to the group mechanism; compared with grouped convolution, MSGC benefits from the learnability, the core of channel pruning, for constructing its group topology, leading to better channel division. The middle spectrum area is unfolded along four dimensions: group-wise, layer-wise, sample-wise, and attention-wise, making it possible to reveal more powerful and interpretable structures. As a result, the proposed module acts as a booster that can reduce the computational cost of the host backbones for general image recognition with even improved predictive accuracy. For example, in the experiments on ImageNet dataset for image classification, MSGC can reduce the multiply-accumulates (MACs) of ResNet-18 and ResNet-50 by half but still increase the Top-1 accuracy by more than 1%. With 35% reduction of MACs, MSGC can also increase the Top-1 accuracy of the MobileNetV2 backbone. Results on MS COCO dataset for object detection show similar observations. Our code and trained models are available at https://github.com/hellozhuo/msgc.

Xiaozheng Zheng, Zhuo Su, Chao Wen et al.

To bridge the physical and virtual worlds for rapidly developed VR/AR applications, the ability to realistically drive 3D full-body avatars is of great significance. Although real-time body tracking with only the head-mounted displays (HMDs) and hand controllers is heavily under-constrained, a carefully designed end-to-end neural network is of great potential to solve the problem by learning from large-scale motion data. To this end, we propose a two-stage framework that can obtain accurate and smooth full-body motions with the three tracking signals of head and hands only. Our framework explicitly models the joint-level features in the first stage and utilizes them as spatiotemporal tokens for alternating spatial and temporal transformer blocks to capture joint-level correlations in the second stage. Furthermore, we design a set of loss terms to constrain the task of a high degree of freedom, such that we can exploit the potential of our joint-level modeling. With extensive experiments on the AMASS motion dataset and real-captured data, we validate the effectiveness of our designs and show our proposed method can achieve more accurate and smooth motion compared to existing approaches.

Yuheng Jiang, Kaixin Yao, Zhuo Su et al.

Convenient 4D modeling of human-object interactions is essential for numerous applications. However, monocular tracking and rendering of complex interaction scenarios remain challenging. In this paper, we propose Instant-NVR, a neural approach for instant volumetric human-object tracking and rendering using a single RGBD camera. It bridges traditional non-rigid tracking with recent instant radiance field techniques via a multi-thread tracking-rendering mechanism. In the tracking front-end, we adopt a robust human-object capture scheme to provide sufficient motion priors. We further introduce a separated instant neural representation with a novel hybrid deformation module for the interacting scene. We also provide an on-the-fly reconstruction scheme of the dynamic/static radiance fields via efficient motion-prior searching. Moreover, we introduce an online key frame selection scheme and a rendering-aware refinement strategy to significantly improve the appearance details for online novel-view synthesis. Extensive experiments demonstrate the effectiveness and efficiency of our approach for the instant generation of human-object radiance fields on the fly, notably achieving real-time photo-realistic novel view synthesis under complex human-object interactions.

Jiehua Zhang, Xueyang Zhang, Zhuo Su et al.

Developing lightweight Deep Convolutional Neural Networks (DCNNs) and Vision Transformers (ViTs) has become one of the focuses in vision research since the low computational cost is essential for deploying vision models on edge devices. Recently, researchers have explored highly computational efficient Binary Neural Networks (BNNs) by binarizing weights and activations of Full-precision Neural Networks. However, the binarization process leads to an enormous accuracy gap between BNN and its full-precision version. One of the primary reasons is that the Sign function with predefined or learned static thresholds limits the representation capacity of binarized architectures since single-threshold binarization fails to utilize activation distributions. To overcome this issue, we introduce the statistics of channel information into explicit thresholds learning for the Sign Function dubbed DySign to generate various thresholds based on input distribution. Our DySign is a straightforward method to reduce information loss and boost the representative capacity of BNNs, which can be flexibly applied to both DCNNs and ViTs (i.e., DyBCNN and DyBinaryCCT) to achieve promising performance improvement. As shown in our extensive experiments. For DCNNs, DyBCNNs based on two backbones (MobileNetV1 and ResNet18) achieve 71.2% and 67.4% top1-accuracy on ImageNet dataset, outperforming baselines by a large margin (i.e., 1.8% and 1.5% respectively). For ViTs, DyBinaryCCT presents the superiority of the convolutional embedding layer in fully binarized ViTs and achieves 56.1% on the ImageNet dataset, which is nearly 9% higher than the baseline.

Jiehua Zhang, Zhuo Su, Li Liu

Face recognition is one of the most active tasks in computer vision and has been widely used in the real world. With great advances made in convolutional neural networks (CNN), lots of face recognition algorithms have achieved high accuracy on various face datasets. However, existing face recognition algorithms based on CNNs are vulnerable to noise. Noise corrupted image patterns could lead to false activations, significantly decreasing face recognition accuracy in noisy situations. To equip CNNs with built-in robustness to noise of different levels, we proposed a Median Pixel Difference Convolutional Network (MeDiNet) by replacing some traditional convolutional layers with the proposed novel Median Pixel Difference Convolutional Layer (MeDiConv) layer. The proposed MeDiNet integrates the idea of traditional multiscale median filtering with deep CNNs. The MeDiNet is tested on the four face datasets (LFW, CA-LFW, CP-LFW, and YTF) with versatile settings on blur kernels, noise intensities, scales, and JPEG quality factors. Extensive experiments show that our MeDiNet can effectively remove noisy pixels in the feature map and suppress the negative impact of noise, leading to achieving limited accuracy loss under these practical noises compared with the standard CNN under clean conditions.

Zhuo Su, Matti Pietikäinen, Li Liu

LBP is a successful hand-crafted feature descriptor in computer vision. However, in the deep learning era, deep neural networks, especially convolutional neural networks (CNNs) can automatically learn powerful task-aware features that are more discriminative and of higher representational capacity. To some extent, such hand-crafted features can be safely ignored when designing deep computer vision models. Nevertheless, due to LBP's preferable properties in visual representation learning, an interesting topic has arisen to explore the value of LBP in enhancing modern deep models in terms of efficiency, memory consumption, and predictive performance. In this paper, we provide a comprehensive review on such efforts which aims to incorporate the LBP mechanism into the design of CNN modules to make deep models stronger. In retrospect of what has been achieved so far, the paper discusses open challenges and directions for future research.

Peng Dai, Yang Zhang, Tao Liu et al.

It is especially challenging to achieve real-time human motion tracking on a standalone VR Head-Mounted Display (HMD) such as Meta Quest and PICO. In this paper, we propose HMD-Poser, the first unified approach to recover full-body motions using scalable sparse observations from HMD and body-worn IMUs. In particular, it can support a variety of input scenarios, such as HMD, HMD+2IMUs, HMD+3IMUs, etc. The scalability of inputs may accommodate users' choices for both high tracking accuracy and easy-to-wear. A lightweight temporal-spatial feature learning network is proposed in HMD-Poser to guarantee that the model runs in real-time on HMDs. Furthermore, HMD-Poser presents online body shape estimation to improve the position accuracy of body joints. Extensive experimental results on the challenging AMASS dataset show that HMD-Poser achieves new state-of-the-art results in both accuracy and real-time performance. We also build a new free-dancing motion dataset to evaluate HMD-Poser's on-device performance and investigate the performance gap between synthetic data and real-captured sensor data. Finally, we demonstrate our HMD-Poser with a real-time Avatar-driving application on a commercial HMD. Our code and free-dancing motion dataset are available https://pico-ai-team.github.io/hmd-poser

Jiancheng Wang, Lidan Liang, Yong Wang et al.

Visual language model (VLM) is rapidly being integrated into safety-critical systems such as autonomous driving, making it an important attack surface for potential backdoor attacks. Existing backdoor attacks mainly rely on unimodal, explicit, and easily detectable triggers, making it difficult to construct both covert and stable attack channels in autonomous driving scenarios. GLA introduces two naturalistic triggers: graffiti-based visual patterns generated via stable diffusion inpainting, which seamlessly blend into urban scenes, and cross-language text triggers, which introduce distributional shifts while maintaining semantic consistency to build robust language-side trigger signals. Experiments on DriveVLM show that GLA requires only a 10\% poisoning ratio to achieve a 90\% Attack Success Rate (ASR) and a 0\% False Positive Rate (FPR). More insidiously, the backdoor does not weaken the model on clean tasks, but instead improves metrics such as BLEU-1, making it difficult for traditional performance-degradation-based detection methods to identify the attack. This study reveals underestimated security threats in self-driving VLMs and provides a new attack paradigm for backdoor evaluation in safety-critical multimodal systems.

Tianmeng Fang, Yong Wang, Zetai Kong et al.

Multimodal large language models have become an important infrastructure for unified processing of visual and linguistic tasks. However, such models are highly susceptible to backdoor implantation during supervised fine-tuning and will steadily output the attacker's predefined harmful responses once a specific trigger pattern is activated. The core challenge of backdoor defense lies in suppressing attack success under low poisoning ratios while preserving the model's normal generation ability. These two objectives are inherently conflicting. Strong suppression often degrades benign performance, whereas weak regularization fails to mitigate backdoor behaviors. To this end, we propose a unified defense framework based on patch augmentation and cross-view regularity, which simultaneously constrains the model's anomalous behaviors in response to triggered patterns from both the feature representation and output distribution levels. Specifically, patch-level data augmentation is combined with cross-view output difference regularization to exploit the fact that backdoor responses are abnormally invariant to non-semantic perturbations and to proactively pull apart the output distributions of the original and perturbed views, thereby significantly suppressing the success rate of backdoor triggering. At the same time, we avoid over-suppression of the model during defense by imposing output entropy constraints, ensuring the quality of normal command generation. Experimental results across three models, two tasks, and six attacks show that our proposed defense method effectively reduces the attack success rate while maintaining a high level of normal text generation capability. Our work enables the secure, controlled deployment of large-scale multimodal models in realistic low-frequency poisoning and covert triggering scenarios.

Zhen Fan, Peng Dai, Zhuo Su et al.

Egocentric human pose estimation (HPE) using wearable sensors is essential for VR/AR applications. Most methods rely solely on either egocentric-view images or sparse Inertial Measurement Unit (IMU) signals, leading to inaccuracies due to self-occlusion in images or the sparseness and drift of inertial sensors. Most importantly, the lack of real-world datasets containing both modalities is a major obstacle to progress in this field. To overcome the barrier, we propose EMHI, a multimodal \textbf{E}gocentric human \textbf{M}otion dataset with \textbf{H}ead-Mounted Display (HMD) and body-worn \textbf{I}MUs, with all data collected under the real VR product suite. Specifically, EMHI provides synchronized stereo images from downward-sloping cameras on the headset and IMU data from body-worn sensors, along with pose annotations in SMPL format. This dataset consists of 885 sequences captured by 58 subjects performing 39 actions, totaling about 28.5 hours of recording. We evaluate the annotations by comparing them with optical marker-based SMPL fitting results. To substantiate the reliability of our dataset, we introduce MEPoser, a new baseline method for multimodal egocentric HPE, which employs a multimodal fusion encoder, temporal feature encoder, and MLP-based regression heads. The experiments on EMHI show that MEPoser outperforms existing single-modal methods and demonstrates the value of our dataset in solving the problem of egocentric HPE. We believe the release of EMHI and the method could advance the research of egocentric HPE and expedite the practical implementation of this technology in VR/AR products.

Long Ma, Yuxin Feng, Yan Zhang et al.

Learning-based image dehazing algorithms have shown remarkable success in synthetic domains. However, real image dehazing is still in suspense due to computational resource constraints and the diversity of real-world scenes. Therefore, there is an urgent need for an algorithm that excels in both efficiency and adaptability to address real image dehazing effectively. This work proposes a Compression-and-Adaptation (CoA) computational flow to tackle these challenges from a divide-and-conquer perspective. First, model compression is performed in the synthetic domain to develop a compact dehazing parameter space, satisfying efficiency demands. Then, a bilevel adaptation in the real domain is introduced to be fearless in unknown real environments by aggregating the synthetic dehazing capabilities during the learning process. Leveraging a succinct design free from additional constraints, our CoA exhibits domain-irrelevant stability and model-agnostic flexibility, effectively bridging the model chasm between synthetic and real domains to further improve its practical utility. Extensive evaluations and analyses underscore the approach's superiority and effectiveness. The code is publicly available at https://github.com/fyxnl/COA.

Yuheng Jiang, Zhehao Shen, Penghao Wang et al.

We have recently seen tremendous progress in photo-real human modeling and rendering. Yet, efficiently rendering realistic human performance and integrating it into the rasterization pipeline remains challenging. In this paper, we present HiFi4G, an explicit and compact Gaussian-based approach for high-fidelity human performance rendering from dense footage. Our core intuition is to marry the 3D Gaussian representation with non-rigid tracking, achieving a compact and compression-friendly representation. We first propose a dual-graph mechanism to obtain motion priors, with a coarse deformation graph for effective initialization and a fine-grained Gaussian graph to enforce subsequent constraints. Then, we utilize a 4D Gaussian optimization scheme with adaptive spatial-temporal regularizers to effectively balance the non-rigid prior and Gaussian updating. We also present a companion compression scheme with residual compensation for immersive experiences on various platforms. It achieves a substantial compression rate of approximately 25 times, with less than 2MB of storage per frame. Extensive experiments demonstrate the effectiveness of our approach, which significantly outperforms existing approaches in terms of optimization speed, rendering quality, and storage overhead.

Huali Xu, Li Liu, Shuaifeng Zhi et al.

Existing Cross-Domain Few-Shot Learning (CDFSL) methods require access to source domain data to train a model in the pre-training phase. However, due to increasing concerns about data privacy and the desire to reduce data transmission and training costs, it is necessary to develop a CDFSL solution without accessing source data. For this reason, this paper explores a Source-Free CDFSL (SF-CDFSL) problem, in which CDFSL is addressed through the use of existing pretrained models instead of training a model with source data, avoiding accessing source data. This paper proposes an Enhanced Information Maximization with Distance-Aware Contrastive Learning (IM-DCL) method to address these challenges. Firstly, we introduce the transductive mechanism for learning the query set. Secondly, information maximization (IM) is explored to map target samples into both individual certainty and global diversity predictions, helping the source model better fit the target data distribution. However, IM fails to learn the decision boundary of the target task. This motivates us to introduce a novel approach called Distance-Aware Contrastive Learning (DCL), in which we consider the entire feature set as both positive and negative sets, akin to Schrodinger's concept of a dual state. Instead of a rigid separation between positive and negative sets, we employ a weighted distance calculation among features to establish a soft classification of the positive and negative sets for the entire feature set. Furthermore, we address issues related to IM by incorporating contrastive constraints between object features and their corresponding positive and negative sets. Evaluations of the 4 datasets in the BSCD-FSL benchmark indicate that the proposed IM-DCL, without accessing the source domain, demonstrates superiority over existing methods, especially in the distant domain task.

Muxin Zhang, Qiao Feng, Zhuo Su et al.

3D human generation is increasingly significant in various applications. However, the direct use of 2D generative methods in 3D generation often results in losing local details, while methods that reconstruct geometry from generated images struggle with global view consistency. In this work, we introduce Joint2Human, a novel method that leverages 2D diffusion models to generate detailed 3D human geometry directly, ensuring both global structure and local details. To achieve this, we employ the Fourier occupancy field (FOF) representation, enabling the direct generation of 3D shapes as preliminary results with 2D generative models. With the proposed high-frequency enhancer and the multi-view recarving strategy, our method can seamlessly integrate the details from different views into a uniform global shape. To better utilize the 3D human prior and enhance control over the generated geometry, we introduce a compact spherical embedding of 3D joints. This allows for an effective guidance of pose during the generation process. Additionally, our method can generate 3D humans guided by textual inputs. Our experimental results demonstrate the capability of our method to ensure global structure, local details, high resolution, and low computational cost simultaneously. More results and the code can be found on our project page at http://cic.tju.edu.cn/faculty/likun/projects/Joint2Human.

Songpengcheng Xia, Yu Zhang, Zhuo Su et al.

Estimating full-body motion using the tracking signals of head and hands from VR devices holds great potential for various applications. However, the sparsity and unique distribution of observations present a significant challenge, resulting in an ill-posed problem with multiple feasible solutions (i.e., hypotheses). This amplifies uncertainty and ambiguity in full-body motion estimation, especially for the lower-body joints. Therefore, we propose a new method, EnvPoser, that employs a two-stage framework to perform full-body motion estimation using sparse tracking signals and pre-scanned environment from VR devices. EnvPoser models the multi-hypothesis nature of human motion through an uncertainty-aware estimation module in the first stage. In the second stage, we refine these multi-hypothesis estimates by integrating semantic and geometric environmental constraints, ensuring that the final motion estimation aligns realistically with both the environmental context and physical interactions. Qualitative and quantitative experiments on two public datasets demonstrate that our method achieves state-of-the-art performance, highlighting significant improvements in human motion estimation within motion-environment interaction scenarios.

Yuheng Jiang, Suyi Jiang, Guoxing Sun et al.

4D modeling of human-object interactions is critical for numerous applications. However, efficient volumetric capture and rendering of complex interaction scenarios, especially from sparse inputs, remain challenging. In this paper, we propose NeuralHOFusion, a neural approach for volumetric human-object capture and rendering using sparse consumer RGBD sensors. It marries traditional non-rigid fusion with recent neural implicit modeling and blending advances, where the captured humans and objects are layerwise disentangled. For geometry modeling, we propose a neural implicit inference scheme with non-rigid key-volume fusion, as well as a template-aid robust object tracking pipeline. Our scheme enables detailed and complete geometry generation under complex interactions and occlusions. Moreover, we introduce a layer-wise human-object texture rendering scheme, which combines volumetric and image-based rendering in both spatial and temporal domains to obtain photo-realistic results. Extensive experiments demonstrate the effectiveness and efficiency of our approach in synthesizing photo-realistic free-view results under complex human-object interactions.

Jiehua Zhang, Zhuo Su, Yanghe Feng et al.

Binary neural networks (BNNs) constrain weights and activations to +1 or -1 with limited storage and computational cost, which is hardware-friendly for portable devices. Recently, BNNs have achieved remarkable progress and been adopted into various fields. However, the performance of BNNs is sensitive to activation distribution. The existing BNNs utilized the Sign function with predefined or learned static thresholds to binarize activations. This process limits representation capacity of BNNs since different samples may adapt to unequal thresholds. To address this problem, we propose a dynamic BNN (DyBNN) incorporating dynamic learnable channel-wise thresholds of Sign function and shift parameters of PReLU. The method aggregates the global information into the hyper function and effectively increases the feature expression ability. The experimental results prove that our method is an effective and straightforward way to reduce information loss and enhance performance of BNNs. The DyBNN based on two backbones of ReActNet (MobileNetV1 and ResNet18) achieve 71.2% and 67.4% top1-accuracy on ImageNet dataset, outperforming baselines by a large margin (i.e., 1.8% and 1.5% respectively).