Xiao Bai

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.

Huang Zhang, Changshuo Wang, Shengwei Tian et al.

In recent years, point cloud representation has become one of the research hotspots in the field of computer vision, and has been widely used in many fields, such as autonomous driving, virtual reality, robotics, etc. Although deep learning techniques have achieved great success in processing regular structured 2D grid image data, there are still great challenges in processing irregular, unstructured point cloud data. Point cloud classification is the basis of point cloud analysis, and many deep learning-based methods have been widely used in this task. Therefore, the purpose of this paper is to provide researchers in this field with the latest research progress and future trends. First, we introduce point cloud acquisition, characteristics, and challenges. Second, we review 3D data representations, storage formats, and commonly used datasets for point cloud classification. We then summarize deep learning-based methods for point cloud classification and complement recent research work. Next, we compare and analyze the performance of the main methods. Finally, we discuss some challenges and future directions for point cloud classification.

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.

Wenjun Miao, Guansong Pang, Trong-Tung Nguyen et al.

Class incremental learning (CIL) aims to learn a model that can not only incrementally accommodate new classes, but also maintain the learned knowledge of old classes. Out-of-distribution (OOD) detection in CIL is to retain this incremental learning ability, while being able to reject unknown samples that are drawn from different distributions of the learned classes. This capability is crucial to the safety of deploying CIL models in open worlds. However, despite remarkable advancements in the respective CIL and OOD detection, there lacks a systematic and large-scale benchmark to assess the capability of advanced CIL models in detecting OOD samples. To fill this gap, in this study we design a comprehensive empirical study to establish such a benchmark, named $\textbf{OpenCIL}$. To this end, we propose two principled frameworks for enabling four representative CIL models with 15 diverse OOD detection methods, resulting in 60 baseline models for OOD detection in CIL. The empirical evaluation is performed on two popular CIL datasets with six commonly-used OOD datasets. One key observation we find through our comprehensive evaluation is that the CIL models can be severely biased towards the OOD samples and newly added classes when they are exposed to open environments. Motivated by this, we further propose a new baseline for OOD detection in CIL, namely Bi-directional Energy Regularization ($\textbf{BER}$), which is specially designed to mitigate these two biases in different CIL models by having energy regularization on both old and new classes. Its superior performance is justified in our experiments. All codes and datasets are open-source at https://github.com/mala-lab/OpenCIL.

Tianqi Li, Guansong Pang, Xiao Bai et al.

Zero-Shot Learning (ZSL) focuses on classifying samples of unseen classes with only their side semantic information presented during training. It cannot handle real-life, open-world scenarios where there are test samples of unknown classes for which neither samples (e.g., images) nor their side semantic information is known during training. Open-Set Recognition (OSR) is dedicated to addressing the unknown class issue, but existing OSR methods are not designed to model the semantic information of the unseen classes. To tackle this combined ZSL and OSR problem, we consider the case of "Zero-Shot Open-Set Recognition" (ZS-OSR), where a model is trained under the ZSL setting but it is required to accurately classify samples from the unseen classes while being able to reject samples from the unknown classes during inference. We perform large experiments on combining existing state-of-the-art ZSL and OSR models for the ZS-OSR task on four widely used datasets adapted from the ZSL task, and reveal that ZS-OSR is a non-trivial task as the simply combined solutions perform badly in distinguishing the unseen-class and unknown-class samples. We further introduce a novel approach specifically designed for ZS-OSR, in which our model learns to generate adversarial semantic embeddings of the unknown classes to train an unknowns-informed ZS-OSR classifier. Extensive empirical results show that our method 1) substantially outperforms the combined solutions in detecting the unknown classes while retaining the classification accuracy on the unseen classes and 2) achieves similar superiority under generalized ZS-OSR settings.

Ruohuan Fang, Guansong Pang, Lei Zhou et al.

Open-World Object Detection (OWOD) extends object detection problem to a realistic and dynamic scenario, where a detection model is required to be capable of detecting both known and unknown objects and incrementally learning newly introduced knowledge. Current OWOD models, such as ORE and OW-DETR, focus on pseudo-labeling regions with high objectness scores as unknowns, whose performance relies heavily on the supervision of known objects. While they can detect the unknowns that exhibit similar features to the known objects, they suffer from a severe label bias problem that they tend to detect all regions (including unknown object regions) that are dissimilar to the known objects as part of the background. To eliminate the label bias, this paper proposes a novel approach that learns an unsupervised discriminative model to recognize true unknown objects from raw pseudo labels generated by unsupervised region proposal methods. The resulting model can be further refined by a classification-free self-training method which iteratively extends pseudo unknown objects to the unlabeled regions. Experimental results show that our method 1) significantly outperforms the prior SOTA in detecting unknown objects while maintaining competitive performance of detecting known object classes on the MS COCO dataset, and 2) achieves better generalization ability on the LVIS and Objects365 datasets.

Zihan Wang, Guansong Pang, Wenjun Miao et al.

Recent advances in Large Visual Language Models (LVLMs) have demonstrated impressive performance across various vision-language tasks by leveraging large-scale image-text pretraining and instruction tuning. However, the security vulnerabilities of LVLMs have become increasingly concerning, particularly their susceptibility to backdoor attacks. Existing backdoor attacks focus on single-target attacks, i.e., targeting a single malicious output associated with a specific trigger. In this work, we uncover multi-target backdoor attacks, where multiple independent triggers corresponding to different attack targets are added in a single pass of training, posing a greater threat to LVLMs in real-world applications. Executing such attacks in LVLMs is challenging since there can be many incorrect trigger-target mappings due to severe feature interference among different triggers. To address this challenge, we propose MTAttack, the first multi-target backdoor attack framework for enforcing accurate multiple trigger-target mappings in LVLMs. The core of MTAttack is a novel optimization method with two constraints, namely Proxy Space Partitioning constraint and Trigger Prototype Anchoring constraint. It jointly optimizes multiple triggers in the latent space, with each trigger independently mapping clean images to a unique proxy class while at the same time guaranteeing their separability. Experiments on popular benchmarks demonstrate a high success rate of MTAttack for multi-target attacks, substantially outperforming existing attack methods. Furthermore, our attack exhibits strong generalizability across datasets and robustness against backdoor defense strategies. These findings highlight the vulnerability of LVLMs to multi-target backdoor attacks and underscore the urgent need for mitigating such threats. Code is available at https://github.com/mala-lab/MTAttack.

Yuchen Wu, Jiahe Li, Fabio Tosi et al.

We present FoundationSLAM, a learning-based monocular dense SLAM system that addresses the absence of geometric consistency in previous flow-based approaches for accurate and robust tracking and mapping. Our core idea is to bridge flow estimation with geometric reasoning by leveraging the guidance from foundation depth models. To this end, we first develop a Hybrid Flow Network that produces geometry-aware correspondences, enabling consistent depth and pose inference across diverse keyframes. To enforce global consistency, we propose a Bi-Consistent Bundle Adjustment Layer that jointly optimizes keyframe pose and depth under multi-view constraints. Furthermore, we introduce a Reliability-Aware Refinement mechanism that dynamically adapts the flow update process by distinguishing between reliable and uncertain regions, forming a closed feedback loop between matching and optimization. Extensive experiments demonstrate that FoundationSLAM achieves superior trajectory accuracy and dense reconstruction quality across multiple challenging datasets, while running in real-time at 18 FPS, demonstrating strong generalization to various scenarios and practical applicability of our method.

Pengcheng Zhang, Xiao Bai, Jin Zheng et al.

End-to-end person search aims to jointly detect and re-identify a target person in raw scene images with a unified model. The detection task unifies all persons while the re-id task discriminates different identities, resulting in conflict optimal objectives. Existing works proposed to decouple end-to-end person search to alleviate such conflict. Yet these methods are still sub-optimal on one or two of the sub-tasks due to their partially decoupled models, which limits the overall person search performance. In this paper, we propose to fully decouple person search towards optimal person search. A task-incremental person search network is proposed to incrementally construct an end-to-end model for the detection and re-id sub-task, which decouples the model architecture for the two sub-tasks. The proposed task-incremental network allows task-incremental training for the two conflicting tasks. This enables independent learning for different objectives thus fully decoupled the model for persons earch. Comprehensive experimental evaluations demonstrate the effectiveness of the proposed fully decoupled models for end-to-end person search.

Tianqi Li, Guansong Pang, Xiao Bai et al.

Existing prompt learning methods have shown certain capabilities in Out-of-Distribution (OOD) detection, but the lack of OOD images in the target dataset in their training can lead to mismatches between OOD images and In-Distribution (ID) categories, resulting in a high false positive rate. To address this issue, we introduce a novel OOD detection method, named 'NegPrompt', to learn a set of negative prompts, each representing a negative connotation of a given class label, for delineating the boundaries between ID and OOD images. It learns such negative prompts with ID data only, without any reliance on external outlier data. Further, current methods assume the availability of samples of all ID classes, rendering them ineffective in open-vocabulary learning scenarios where the inference stage can contain novel ID classes not present during training. In contrast, our learned negative prompts are transferable to novel class labels. Experiments on various ImageNet benchmarks show that NegPrompt surpasses state-of-the-art prompt-learning-based OOD detection methods and maintains a consistent lead in hard OOD detection in closed- and open-vocabulary classification scenarios. Code is available at https://github.com/mala-lab/negprompt.

Wenjun Miao, Guansong Pang, Tianqi Li et al.

Existing out-of-distribution (OOD) methods have shown great success on balanced datasets but become ineffective in long-tailed recognition (LTR) scenarios where 1) OOD samples are often wrongly classified into head classes and/or 2) tail-class samples are treated as OOD samples. To address these issues, current studies fit a prior distribution of auxiliary/pseudo OOD data to the long-tailed in-distribution (ID) data. However, it is difficult to obtain such an accurate prior distribution given the unknowingness of real OOD samples and heavy class imbalance in LTR. A straightforward solution to avoid the requirement of this prior is to learn an outlier class to encapsulate the OOD samples. The main challenge is then to tackle the aforementioned confusion between OOD samples and head/tail-class samples when learning the outlier class. To this end, we introduce a novel calibrated outlier class learning (COCL) approach, in which 1) a debiased large margin learning method is introduced in the outlier class learning to distinguish OOD samples from both head and tail classes in the representation space and 2) an outlier-class-aware logit calibration method is defined to enhance the long-tailed classification confidence. Extensive empirical results on three popular benchmarks CIFAR10-LT, CIFAR100-LT, and ImageNet-LT demonstrate that COCL substantially outperforms state-of-the-art OOD detection methods in LTR while being able to improve the classification accuracy on ID data. Code is available at https://github.com/mala-lab/COCL.

Ruohuan Fang, Guansong Pang, Xiao Bai

Open-Vocabulary Object Detection (OVOD) aims to detect novel objects beyond a given set of base categories on which the detection model is trained. Recent OVOD methods focus on adapting the image-level pre-trained vision-language models (VLMs), such as CLIP, to a region-level object detection task via, eg., region-level knowledge distillation, regional prompt learning, or region-text pre-training, to expand the detection vocabulary. These methods have demonstrated remarkable performance in recognizing regional visual concepts, but they are weak in exploiting the VLMs' powerful global scene understanding ability learned from the billion-scale image-level text descriptions. This limits their capability in detecting hard objects of small, blurred, or occluded appearance from novel/base categories, whose detection heavily relies on contextual information. To address this, we propose a novel approach, namely Simple Image-level Classification for Context-Aware Detection Scoring (SIC-CADS), to leverage the superior global knowledge yielded from CLIP for complementing the current OVOD models from a global perspective. The core of SIC-CADS is a multi-modal multi-label recognition (MLR) module that learns the object co-occurrence-based contextual information from CLIP to recognize all possible object categories in the scene. These image-level MLR scores can then be utilized to refine the instance-level detection scores of the current OVOD models in detecting those hard objects. This is verified by extensive empirical results on two popular benchmarks, OV-LVIS and OV-COCO, which show that SIC-CADS achieves significant and consistent improvement when combined with different types of OVOD models. Further, SIC-CADS also improves the cross-dataset generalization ability on Objects365 and OpenImages. The code is available at https://github.com/mala-lab/SIC-CADS.

Wenjun Miao, Guansong Pang, Jin Zheng et al.

One key challenge in Out-of-Distribution (OOD) detection is the absence of ground-truth OOD samples during training. One principled approach to address this issue is to use samples from external datasets as outliers (i.e., pseudo OOD samples) to train OOD detectors. However, we find empirically that the outlier samples often present a distribution shift compared to the true OOD samples, especially in Long-Tailed Recognition (LTR) scenarios, where ID classes are heavily imbalanced, \ie, the true OOD samples exhibit very different probability distribution to the head and tailed ID classes from the outliers. In this work, we propose a novel approach, namely normalized outlier distribution adaptation (AdaptOD), to tackle this distribution shift problem. One of its key components is dynamic outlier distribution adaptation that effectively adapts a vanilla outlier distribution based on the outlier samples to the true OOD distribution by utilizing the OOD knowledge in the predicted OOD samples during inference. Further, to obtain a more reliable set of predicted OOD samples on long-tailed ID data, a novel dual-normalized energy loss is introduced in AdaptOD, which leverages class- and sample-wise normalized energy to enforce a more balanced prediction energy on imbalanced ID samples. This helps avoid bias toward the head samples and learn a substantially better vanilla outlier distribution than existing energy losses during training. It also eliminates the need of manually tuning the sensitive margin hyperparameters in energy losses. Empirical results on three popular benchmarks for OOD detection in LTR show the superior performance of AdaptOD over state-of-the-art methods. Code is available at https://github.com/mala-lab/AdaptOD.

Pengcheng Zhang, Xiaohan Yu, Xiao Bai et al.

The development of person search techniques has been greatly promoted in recent years for its superior practicality and challenging goals. Despite their significant progress, existing person search models still lack the ability to continually learn from increaseing real-world data and adaptively process input from different domains. To this end, this work introduces the continual person search task that sequentially learns on multiple domains and then performs person search on all seen domains. This requires balancing the stability and plasticity of the model to continually learn new knowledge without catastrophic forgetting. For this, we propose a Prompt-based Continual Person Search (PoPS) model in this paper. First, we design a compositional person search transformer to construct an effective pre-trained transformer without exhaustive pre-training from scratch on large-scale person search data. This serves as the fundamental for prompt-based continual learning. On top of that, we design a domain incremental prompt pool with a diverse attribute matching module. For each domain, we independently learn a set of prompts to encode the domain-oriented knowledge. Meanwhile, we jointly learn a group of diverse attribute projections and prototype embeddings to capture discriminative domain attributes. By matching an input image with the learned attributes across domains, the learned prompts can be properly selected for model inference. Extensive experiments are conducted to validate the proposed method for continual person search. The source code is available at https://github.com/PatrickZad/PoPS.

Yuchen Wu, Jiahe Li, Xiaohan Yu et al.

Monocular visual SLAM enables 3D reconstruction from internet video and autonomous navigation on resource-constrained platforms, yet suffers from scale drift, i.e., the gradual divergence of estimated scale over long sequences. Existing frame-to-frame methods achieve real-time performance through local optimization but accumulate scale drift due to the lack of global constraints among independent windows. To address this, we propose SCE-SLAM, an end-to-end SLAM system that maintains scale consistency through scene coordinate embeddings, which are learned patch-level representations encoding 3D geometric relationships under a canonical scale reference. The framework consists of two key modules: geometry-guided aggregation that leverages 3D spatial proximity to propagate scale information from historical observations through geometry-modulated attention, and scene coordinate bundle adjustment that anchors current estimates to the reference scale through explicit 3D coordinate constraints decoded from the scene coordinate embeddings. Experiments on KITTI, Waymo, and vKITTI demonstrate substantial improvements: our method reduces absolute trajectory error by 8.36m on KITTI compared to the best prior approach, while maintaining 36 FPS and achieving scale consistency across large-scale scenes.

Jiahe Li, Jiawei Zhang, Youmin Zhang et al.

Reconstructing accurate surfaces with radiance fields has achieved remarkable progress in recent years. However, prevailing approaches, primarily based on Gaussian Splatting, are increasingly constrained by representational bottlenecks. In this paper, we introduce GeoSVR, an explicit voxel-based framework that explores and extends the under-investigated potential of sparse voxels for achieving accurate, detailed, and complete surface reconstruction. As strengths, sparse voxels support preserving the coverage completeness and geometric clarity, while corresponding challenges also arise from absent scene constraints and locality in surface refinement. To ensure correct scene convergence, we first propose a Voxel-Uncertainty Depth Constraint that maximizes the effect of monocular depth cues while presenting a voxel-oriented uncertainty to avoid quality degradation, enabling effective and robust scene constraints yet preserving highly accurate geometries. Subsequently, Sparse Voxel Surface Regularization is designed to enhance geometric consistency for tiny voxels and facilitate the voxel-based formation of sharp and accurate surfaces. Extensive experiments demonstrate our superior performance compared to existing methods across diverse challenging scenarios, excelling in geometric accuracy, detail preservation, and reconstruction completeness while maintaining high efficiency. Code is available at https://github.com/Fictionarry/GeoSVR.

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.

Jiahe Li, Jiawei Zhang, Xiao Bai et al.

Surface reconstruction with differentiable rendering has achieved impressive performance in recent years, yet the pervasive photometric ambiguities have strictly bottlenecked existing approaches. This paper presents AmbiSuR, a framework that explores an intrinsic solution upon Gaussian Splatting for the photometric ambiguity-robust surface 3D reconstruction with high performance. Starting by revisiting the foundation, our investigation uncovers two built-in primitive-wise ambiguities in representation, while revealing an intrinsic potential for ambiguity self-indication in Gaussian Splatting. Stemming from these, a photometric disambiguation is first introduced, constraining ill-posed geometry solution for definite surface formation. Then, we propose an ambiguity indication module that unleashes the self-indication potential to identify and further guide correcting underconstrained reconstructions. Extensive experiments demonstrate our superior surface reconstructions compared to existing methods across various challenging scenarios, excelling in broad compatibility. Project: https://fictionarry.github.io/AmbiSuR-Proj/ .

Peng Di, Faqiang Chen, Xiao Bai et al.

The escalating complexity of modern software imposes an unsustainable operational burden on Site Reliability Engineering (SRE) teams, demanding AI-driven automation that can emulate expert diagnostic reasoning. Existing solutions, from traditional AI methods to general-purpose multi-agent systems, fall short: they either lack deep causal reasoning or are not tailored for the specialized, investigative workflows unique to SRE. To address this gap, we present OpenDerisk, a specialized, open-source multi-agent framework architected for SRE. OpenDerisk integrates a diagnostic-native collaboration model, a pluggable reasoning engine, a knowledge engine, and a standardized protocol (MCP) to enable specialist agents to collectively solve complex, multi-domain problems. Our comprehensive evaluation demonstrates that OpenDerisk significantly outperforms state-of-the-art baselines in both accuracy and efficiency. This effectiveness is validated by its large-scale production deployment at Ant Group, where it serves over 3,000 daily users across diverse scenarios, confirming its industrial-grade scalability and practical impact. OpenDerisk is open source and available at https://github.com/derisk-ai/OpenDerisk/

Yang Liu, Lei Zhou, Xiao Bai et al.

Zero-shot learning (ZSL) aims to recognize novel classes by transferring semantic knowledge from seen classes to unseen classes. Since semantic knowledge is built on attributes shared between different classes, which are highly local, strong prior for localization of object attribute is beneficial for visual-semantic embedding. Interestingly, when recognizing unseen images, human would also automatically gaze at regions with certain semantic clue. Therefore, we introduce a novel goal-oriented gaze estimation module (GEM) to improve the discriminative attribute localization based on the class-level attributes for ZSL. We aim to predict the actual human gaze location to get the visual attention regions for recognizing a novel object guided by attribute description. Specifically, the task-dependent attention is learned with the goal-oriented GEM, and the global image features are simultaneously optimized with the regression of local attribute features. Experiments on three ZSL benchmarks, i.e., CUB, SUN and AWA2, show the superiority or competitiveness of our proposed method against the state-of-the-art ZSL methods. The ablation analysis on real gaze data CUB-VWSW also validates the benefits and accuracy of our gaze estimation module. This work implies the promising benefits of collecting human gaze dataset and automatic gaze estimation algorithms on high-level computer vision tasks. The code is available at https://github.com/osierboy/GEM-ZSL.

Yang Liu, Lei Zhou, Xiao Bai et al.

Zero-shot learning (ZSL) aims to recognize novel classes by transferring semantic knowledge from seen classes to unseen classes. Though many ZSL methods rely on a direct mapping between the visual and the semantic space, the calibration deviation and hubness problem limit the generalization capability to unseen classes. Recently emerged generative ZSL methods generate unseen image features to transform ZSL into a supervised classification problem. However, most generative models still suffer from the seen-unseen bias problem as only seen data is used for training. To address these issues, we propose a novel bidirectional embedding based generative model with a tight visual-semantic coupling constraint. We learn a unified latent space that calibrates the embedded parametric distributions of both visual and semantic spaces. Since the embedding from high-dimensional visual features comprise much non-semantic information, the alignment of visual and semantic in latent space would inevitably been deviated. Therefore, we introduce information bottleneck (IB) constraint to ZSL for the first time to preserve essential attribute information during the mapping. Specifically, we utilize the uncertainty estimation and the wake-sleep procedure to alleviate the feature noises and improve model abstraction capability. In addition, our method can be easily extended to transductive ZSL setting by generating labels for unseen images. We then introduce a robust loss to solve this label noise problem. Extensive experimental results show that our method outperforms the state-of-the-art methods in different ZSL settings on most benchmark datasets. The code will be available at https://github.com/osierboy/IBZSL.

Youmin Zhang, Yimin Chen, Xiao Bai et al.

State-of-the-art deep learning based stereo matching approaches treat disparity estimation as a regression problem, where loss function is directly defined on true disparities and their estimated ones. However, disparity is just a byproduct of a matching process modeled by cost volume, while indirectly learning cost volume driven by disparity regression is prone to overfitting since the cost volume is under constrained. In this paper, we propose to directly add constraints to the cost volume by filtering cost volume with unimodal distribution peaked at true disparities. In addition, variances of the unimodal distributions for each pixel are estimated to explicitly model matching uncertainty under different contexts. The proposed architecture achieves state-of-the-art performance on Scene Flow and two KITTI stereo benchmarks. In particular, our method ranked the $1^{st}$ place of KITTI 2012 evaluation and the $4^{th}$ place of KITTI 2015 evaluation (recorded on 2019.8.20). The codes of AcfNet are available at: https://github.com/DeepMotionAIResearch/DenseMatchingBenchmark.

Jiawei Zhang, Jiahe Li, Xiaohan Yu et al.

3D Gaussian Splatting (3DGS) creates a radiance field consisting of 3D Gaussians to represent a scene. With sparse training views, 3DGS easily suffers from overfitting, negatively impacting rendering. This paper introduces a new co-regularization perspective for improving sparse-view 3DGS. When training two 3D Gaussian radiance fields, we observe that the two radiance fields exhibit point disagreement and rendering disagreement that can unsupervisedly predict reconstruction quality, stemming from the randomness of densification implementation. We further quantify the two disagreements and demonstrate the negative correlation between them and accurate reconstruction, which allows us to identify inaccurate reconstruction without accessing ground-truth information. Based on the study, we propose CoR-GS, which identifies and suppresses inaccurate reconstruction based on the two disagreements: (1) Co-pruning considers Gaussians that exhibit high point disagreement in inaccurate positions and prunes them. (2) Pseudo-view co-regularization considers pixels that exhibit high rendering disagreement are inaccurate and suppress the disagreement. Results on LLFF, Mip-NeRF360, DTU, and Blender demonstrate that CoR-GS effectively regularizes the scene geometry, reconstructs the compact representations, and achieves state-of-the-art novel view synthesis quality under sparse training views.

Jiahe Li, Jiawei Zhang, Xiao Bai et al.

Radiance fields have demonstrated impressive performance in synthesizing lifelike 3D talking heads. However, due to the difficulty in fitting steep appearance changes, the prevailing paradigm that presents facial motions by directly modifying point appearance may lead to distortions in dynamic regions. To tackle this challenge, we introduce TalkingGaussian, a deformation-based radiance fields framework for high-fidelity talking head synthesis. Leveraging the point-based Gaussian Splatting, facial motions can be represented in our method by applying smooth and continuous deformations to persistent Gaussian primitives, without requiring to learn the difficult appearance change like previous methods. Due to this simplification, precise facial motions can be synthesized while keeping a highly intact facial feature. Under such a deformation paradigm, we further identify a face-mouth motion inconsistency that would affect the learning of detailed speaking motions. To address this conflict, we decompose the model into two branches separately for the face and inside mouth areas, therefore simplifying the learning tasks to help reconstruct more accurate motion and structure of the mouth region. Extensive experiments demonstrate that our method renders high-quality lip-synchronized talking head videos, with better facial fidelity and higher efficiency compared with previous methods.

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.

Jiahe Li, Jiawei Zhang, Xiao Bai et al.

Despite exhibiting impressive performance in synthesizing lifelike personalized 3D talking heads, prevailing methods based on radiance fields suffer from high demands for training data and time for each new identity. This paper introduces InsTaG, a 3D talking head synthesis framework that allows a fast learning of realistic personalized 3D talking head from few training data. Built upon a lightweight 3DGS person-specific synthesizer with universal motion priors, InsTaG achieves high-quality and fast adaptation while preserving high-level personalization and efficiency. As preparation, we first propose an Identity-Free Pre-training strategy that enables the pre-training of the person-specific model and encourages the collection of universal motion priors from long-video data corpus. To fully exploit the universal motion priors to learn an unseen new identity, we then present a Motion-Aligned Adaptation strategy to adaptively align the target head to the pre-trained field, and constrain a robust dynamic head structure under few training data. Experiments demonstrate our outstanding performance and efficiency under various data scenarios to render high-quality personalized talking heads.

Clark Mingxuan Ju, Tong Zhao, Leonardo Neves et al.

Effective item identifiers (IDs) are an important component for recommender systems (RecSys) in practice, and are commonly adopted in many use cases such as retrieval and ranking. IDs can encode collaborative filtering signals within training data, such that RecSys models can extrapolate during the inference and personalize the prediction based on users' behavioral histories. Recently, Semantic IDs (SIDs) have become a trending paradigm for RecSys. In comparison to the conventional atomic ID, an SID is an ordered list of codes, derived from tokenizers such as residual quantization, applied to semantic representations commonly extracted from foundation models or collaborative signals. SIDs have drastically smaller cardinality than the atomic counterpart, and induce semantic clustering in the ID space. At Snapchat, we apply SIDs as auxiliary features for ranking models, and also explore SIDs as additional retrieval sources in different ML applications. In this paper, we discuss practical technical challenges we encountered while applying SIDs, experiments we have conducted, and design choices we have iterated to mitigate these challenges. Backed by promising offline results on both internal data and academic benchmarks as well as online A/B studies, SID variants have been launched in multiple production models with positive metrics impact.

Meiying Gu, Jiawei Zhang, Jiahe Li et al.

Recent advances in optimizing Gaussian Splatting for scene geometry have enabled efficient reconstruction of detailed surfaces from images. However, when input views are sparse, such optimization is prone to overfitting, leading to suboptimal reconstruction quality. Existing approaches address this challenge by employing flattened Gaussian primitives to better fit surface geometry, combined with depth regularization to alleviate geometric ambiguities under limited viewpoints. Nevertheless, the increased anisotropy inherent in flattened Gaussians exacerbates overfitting in sparse-view scenarios, hindering accurate surface fitting and degrading novel view synthesis performance. In this paper, we propose \net{}, a method that reconstructs more accurate and detailed surfaces while preserving high-quality novel view rendering. Our key insight is to introduce Stereo Geometry-Texture Alignment, which bridges rendering quality and geometry estimation, thereby jointly enhancing both surface reconstruction and view synthesis. In addition, we present a Pseudo-Feature Enhanced Geometry Consistency that enforces multi-view geometric consistency by incorporating both training and unseen views, effectively mitigating overfitting caused by sparse supervision. Extensive experiments on the DTU, BlendedMVS, and Mip-NeRF360 datasets demonstrate that our method achieves the state-of-the-art performance.

Suihanjin Yu, Youmin Zhang, Chen Wang et al.

In optical flow estimation task, coarse-to-fine (C2F) warping strategy is widely used to deal with the large displacement problem and provides efficiency and speed. However, limited by the small search range between the first images and warped second images, current coarse-to-fine optical flow networks fail to capture small and fast-moving objects which disappear at coarse resolution levels. To address this problem, we introduce a lightweight but effective Global Matching Component (GMC) to grab global matching features. We propose a new Hybrid Matching Optical Flow Network (HMFlow) by integrating GMC into existing coarse-to-fine networks seamlessly. Besides keeping in high accuracy and small model size, our proposed HMFlow can apply global matching features to guide the network to discover the small and fast-moving objects mismatched by local matching features. We also build a new dataset, named Small and Fast-Moving Chairs (SFChairs), for evaluation. The experimental results show that our proposed network achieves considerable performance, especially at regions with small and fast-moving objects.

Cheng Yan, Guansong Pang, Xiao Bai et al.

Person Re-IDentification (ReID) aims at re-identifying persons from different viewpoints across multiple cameras. Capturing the fine-grained appearance differences is often the key to accurate person ReID, because many identities can be differentiated only when looking into these fine-grained differences. However, most state-of-the-art person ReID approaches, typically driven by a triplet loss, fail to effectively learn the fine-grained features as they are focused more on differentiating large appearance differences. To address this issue, we introduce a novel pairwise loss function that enables ReID models to learn the fine-grained features by adaptively enforcing an exponential penalization on the images of small differences and a bounded penalization on the images of large differences. The proposed loss is generic and can be used as a plugin to replace the triplet loss to significantly enhance different types of state-of-the-art approaches. Experimental results on four benchmark datasets show that the proposed loss substantially outperforms a number of popular loss functions by large margins; and it also enables significantly improved data efficiency.

Wasi Uddin Ahmad, Xiao Bai, Soomin Lee et al.

Natural language processing techniques have demonstrated promising results in keyphrase generation. However, one of the major challenges in \emph{neural} keyphrase generation is processing long documents using deep neural networks. Generally, documents are truncated before given as inputs to neural networks. Consequently, the models may miss essential points conveyed in the target document. To overcome this limitation, we propose \emph{SEG-Net}, a neural keyphrase generation model that is composed of two major components, (1) a selector that selects the salient sentences in a document and (2) an extractor-generator that jointly extracts and generates keyphrases from the selected sentences. SEG-Net uses Transformer, a self-attentive architecture, as the basic building block with a novel \emph{layer-wise} coverage attention to summarize most of the points discussed in the document. The experimental results on seven keyphrase generation benchmarks from scientific and web documents demonstrate that SEG-Net outperforms the state-of-the-art neural generative methods by a large margin.

Lu Yu, Shichao Pei, Chuxu Zhang et al.

Pairwise ranking models have been widely used to address recommendation problems. The basic idea is to learn the rank of users' preferred items through separating items into \emph{positive} samples if user-item interactions exist, and \emph{negative} samples otherwise. Due to the limited number of observable interactions, pairwise ranking models face serious \emph{class-imbalance} issues. Our theoretical analysis shows that current sampling-based methods cause the vertex-level imbalance problem, which makes the norm of learned item embeddings towards infinite after a certain training iterations, and consequently results in vanishing gradient and affects the model inference results. We thus propose an efficient \emph{\underline{Vi}tal \underline{N}egative \underline{S}ampler} (VINS) to alleviate the class-imbalance issue for pairwise ranking model, in particular for deep learning models optimized by gradient methods. The core of VINS is a bias sampler with reject probability that will tend to accept a negative candidate with a larger degree weight than the given positive item. Evaluation results on several real datasets demonstrate that the proposed sampling method speeds up the training procedure 30\% to 50\% for ranking models ranging from shallow to deep, while maintaining and even improving the quality of ranking results in top-N item recommendation.

Haotong Qin, Ruihao Gong, Xianglong Liu et al.

The binary neural network, largely saving the storage and computation, serves as a promising technique for deploying deep models on resource-limited devices. However, the binarization inevitably causes severe information loss, and even worse, its discontinuity brings difficulty to the optimization of the deep network. To address these issues, a variety of algorithms have been proposed, and achieved satisfying progress in recent years. In this paper, we present a comprehensive survey of these algorithms, mainly categorized into the native solutions directly conducting binarization, and the optimized ones using techniques like minimizing the quantization error, improving the network loss function, and reducing the gradient error. We also investigate other practical aspects of binary neural networks such as the hardware-friendly design and the training tricks. Then, we give the evaluation and discussions on different tasks, including image classification, object detection and semantic segmentation. Finally, the challenges that may be faced in future research are prospected.

Guansong Pang, Cheng Yan, Chunhua Shen et al.

Video anomaly detection is of critical practical importance to a variety of real applications because it allows human attention to be focused on events that are likely to be of interest, in spite of an otherwise overwhelming volume of video. We show that applying self-trained deep ordinal regression to video anomaly detection overcomes two key limitations of existing methods, namely, 1) being highly dependent on manually labeled normal training data; and 2) sub-optimal feature learning. By formulating a surrogate two-class ordinal regression task we devise an end-to-end trainable video anomaly detection approach that enables joint representation learning and anomaly scoring without manually labeled normal/abnormal data. Experiments on eight real-world video scenes show that our proposed method outperforms state-of-the-art methods that require no labeled training data by a substantial margin, and enables easy and accurate localization of the identified anomalies. Furthermore, we demonstrate that our method offers effective human-in-the-loop anomaly detection which can be critical in applications where anomalies are rare and the false-negative cost is high.

Cheng Yan, Guansong Pang, Xiao Bai et al.

Person re-identification (ReID) aims at re-identifying persons from different viewpoints across multiple cameras, of which it is of great importance to learn multifaceted features expressed in different parts of a person, e.g., clothes, bags, and other accessories in the main body, appearance in the head, and shoes in the foot. To learn such features, existing methods are focused on the striping-based approach that builds multi-branch neural networks to learn local features in each part of the identities, with one-branch network dedicated to one part. This results in complex models with a large number of parameters. To address this issue, this paper proposes to learn the multifaceted features in a simple unified single-branch neural network. The Unified Multifaceted Feature Learning (UMFL) framework is introduced to fulfill this goal, which consists of two key collaborative modules: compound batch image erasing (including batch constant erasing and random erasing) and hierarchical structured loss. The loss structures the augmented images resulted by the two types of image erasing in a two-level hierarchy and enforces multifaceted attention to different parts. As we show in the extensive experimental results on four benchmark person ReID datasets, despite the use of significantly simplified network structure, our method performs substantially better than state-of-the-art competing methods. Our method can also effectively generalize to vehicle ReID, achieving similar improvement on two vehicle ReID datasets.

Dong Wang, Lei Zhou, Xiao Bai et al.

Acceleration of convolutional neural network has received increasing attention during the past several years. Among various acceleration techniques, filter pruning has its inherent merit by effectively reducing the number of convolution filters. However, most filter pruning methods resort to tedious and time-consuming layer-by-layer pruning-recovery strategy to avoid a significant drop of accuracy. In this paper, we present an efficient filter pruning framework to solve this problem. Our method accelerates the network in one-step pruning-recovery manner with a novel optimization objective function, which achieves higher accuracy with much less cost compared with existing pruning methods. Furthermore, our method allows network compression with global filter pruning. Given a global pruning rate, it can adaptively determine the pruning rate for each single convolutional layer, while these rates are often set as hyper-parameters in previous approaches. Evaluated on VGG-16 and ResNet-50 using ImageNet, our approach outperforms several state-of-the-art methods with less accuracy drop under the same and even much fewer floating-point operations (FLOPs).

Jelena Gligorijevic, Djordje Gligorijevic, Ivan Stojkovic et al.

In sponsored search it is critical to match ads that are relevant to a query and to accurately predict their likelihood of being clicked. Commercial search engines typically use machine learning models for both query-ad relevance matching and click-through-rate (CTR) prediction. However, matching models are based on the similarity between a query and an ad, ignoring the fact that a retrieved ad may not attract clicks, while click models rely on click history, being of limited use for new queries and ads. We propose a deeply supervised architecture that jointly learns the semantic embeddings of a query and an ad as well as their corresponding CTR.We also propose a novel cohort negative sampling technique for learning implicit negative signals. We trained the proposed architecture using one billion query-ad pairs from a major commercial web search engine. This architecture improves the best-performing baseline deep neural architectures by 2\% of AUC for CTR prediction and by statistically significant 0.5\% of NDCG for query-ad matching.

Dong Wang, Lei Zhou, Xueni Zhang et al.

While the research on convolutional neural networks (CNNs) is progressing quickly, the real-world deployment of these models is often limited by computing resources and memory constraints. In this paper, we address this issue by proposing a novel filter pruning method to compress and accelerate CNNs. Our work is based on the linear relationship identified in different feature map subspaces via visualization of feature maps. Such linear relationship implies that the information in CNNs is redundant. Our method eliminates the redundancy in convolutional filters by applying subspace clustering to feature maps. In this way, most of the representative information in the network can be retained in each cluster. Therefore, our method provides an effective solution to filter pruning for which most existing methods directly remove filters based on simple heuristics. The proposed method is independent of the network structure, thus it can be adopted by any off-the-shelf deep learning libraries. Experiments on different networks and tasks show that our method outperforms existing techniques before fine-tuning, and achieves the state-of-the-art results after fine-tuning.

Lei Zhou, Xiao Bai, Xianglong Liu et al.

Subspace clustering is a useful technique for many computer vision applications in which the intrinsic dimension of high-dimensional data is often smaller than the ambient dimension. Spectral clustering, as one of the main approaches to subspace clustering, often takes on a sparse representation or a low-rank representation to learn a block diagonal self-representation matrix for subspace generation. However, existing methods require solving a large scale convex optimization problem with a large set of data, with computational complexity reaches O(N^3) for N data points. Therefore, the efficiency and scalability of traditional spectral clustering methods can not be guaranteed for large scale datasets. In this paper, we propose a subspace clustering model based on the Kronecker product. Due to the property that the Kronecker product of a block diagonal matrix with any other matrix is still a block diagonal matrix, we can efficiently learn the representation matrix which is formed by the Kronecker product of k smaller matrices. By doing so, our model significantly reduces the computational complexity to O(kN^{3/k}). Furthermore, our model is general in nature, and can be adapted to different regularization based subspace clustering methods. Experimental results on two public datasets show that our model significantly improves the efficiency compared with several state-of-the-art methods. Moreover, we have conducted experiments on synthetic data to verify the scalability of our model for large scale datasets.

Bora Edizel, Amin Mantrach, Xiao Bai

Predicting the click-through rate of an advertisement is a critical component of online advertising platforms. In sponsored search, the click-through rate estimates the probability that a displayed advertisement is clicked by a user after she submits a query to the search engine. Commercial search engines typically rely on machine learning models trained with a large number of features to make such predictions. This is inevitably requires a lot of engineering efforts to define, compute, and select the appropriate features. In this paper, we propose two novel approaches (one working at character level and the other working at word level) that use deep convolutional neural networks to predict the click-through rate of a query-advertisement pair. Specially, the proposed architectures only consider the textual content appearing in a query-advertisement pair as input, and produce as output a click-through rate prediction. By comparing the character-level model with the word-level model, we show that language representation can be learnt from scratch at character level when trained on enough data. Through extensive experiments using billions of query-advertisement pairs of a popular commercial search engine, we demonstrate that both approaches significantly outperform a baseline model built on well-selected text features and a state-of-the-art word2vec-based approach. Finally, by combining the predictions of the deep models introduced in this study with the prediction of the model in production of the same commercial search engine, we significantly improve the accuracy and the calibration of the click-through rate prediction of the production system.

Jie Liang, Jun Zhou, Yuntao Qian et al.

Spectral-spatial processing has been increasingly explored in remote sensing hyperspectral image classification. While extensive studies have focused on developing methods to improve the classification accuracy, experimental setting and design for method evaluation have drawn little attention. In the scope of supervised classification, we find that traditional experimental designs for spectral processing are often improperly used in the spectral-spatial processing context, leading to unfair or biased performance evaluation. This is especially the case when training and testing samples are randomly drawn from the same image - a practice that has been commonly adopted in the experiments. Under such setting, the dependence caused by overlap between the training and testing samples may be artificially enhanced by some spatial information processing methods such as spatial filtering and morphological operation. Such interaction between training and testing sets has violated data independence assumption that is abided by supervised learning theory and performance evaluation mechanism. Therefore, the widely adopted pixel-based random sampling strategy is not always suitable to evaluate spectral-spatial classification algorithms because it is difficult to determine whether the improvement of classification accuracy is caused by incorporating spatial information into classifier or by increasing the overlap between training and testing samples. To partially solve this problem, we propose a novel controlled random sampling strategy for spectral-spatial methods. It can greatly reduce the overlap between training and testing samples and provides more objective and accurate evaluation.