Xuejin Chen

Binxin Yang, Shuyang Gu, Bo Zhang et al. · microsoft-research

Language-guided image editing has achieved great success recently. In this paper, for the first time, we investigate exemplar-guided image editing for more precise control. We achieve this goal by leveraging self-supervised training to disentangle and re-organize the source image and the exemplar. However, the naive approach will cause obvious fusing artifacts. We carefully analyze it and propose an information bottleneck and strong augmentations to avoid the trivial solution of directly copying and pasting the exemplar image. Meanwhile, to ensure the controllability of the editing process, we design an arbitrary shape mask for the exemplar image and leverage the classifier-free guidance to increase the similarity to the exemplar image. The whole framework involves a single forward of the diffusion model without any iterative optimization. We demonstrate that our method achieves an impressive performance and enables controllable editing on in-the-wild images with high fidelity.

Kaizhi Yang, Xiaoshuai Zhang, Zhiao Huang et al.

We present MovingParts, a NeRF-based method for dynamic scene reconstruction and part discovery. We consider motion as an important cue for identifying parts, that all particles on the same part share the common motion pattern. From the perspective of fluid simulation, existing deformation-based methods for dynamic NeRF can be seen as parameterizing the scene motion under the Eulerian view, i.e., focusing on specific locations in space through which the fluid flows as time passes. However, it is intractable to extract the motion of constituting objects or parts using the Eulerian view representation. In this work, we introduce the dual Lagrangian view and enforce representations under the Eulerian/Lagrangian views to be cycle-consistent. Under the Lagrangian view, we parameterize the scene motion by tracking the trajectory of particles on objects. The Lagrangian view makes it convenient to discover parts by factorizing the scene motion as a composition of part-level rigid motions. Experimentally, our method can achieve fast and high-quality dynamic scene reconstruction from even a single moving camera, and the induced part-based representation allows direct applications of part tracking, animation, 3D scene editing, etc.

Kai Cheng, Xiaoxiao Long, Wei Yin et al.

Multi-camera setups find widespread use across various applications, such as autonomous driving, as they greatly expand sensing capabilities. Despite the fast development of Neural radiance field (NeRF) techniques and their wide applications in both indoor and outdoor scenes, applying NeRF to multi-camera systems remains very challenging. This is primarily due to the inherent under-calibration issues in multi-camera setup, including inconsistent imaging effects stemming from separately calibrated image signal processing units in diverse cameras, and system errors arising from mechanical vibrations during driving that affect relative camera poses. In this paper, we present UC-NeRF, a novel method tailored for novel view synthesis in under-calibrated multi-view camera systems. Firstly, we propose a layer-based color correction to rectify the color inconsistency in different image regions. Second, we propose virtual warping to generate more viewpoint-diverse but color-consistent virtual views for color correction and 3D recovery. Finally, a spatiotemporally constrained pose refinement is designed for more robust and accurate pose calibration in multi-camera systems. Our method not only achieves state-of-the-art performance of novel view synthesis in multi-camera setups, but also effectively facilitates depth estimation in large-scale outdoor scenes with the synthesized novel views.

Qingyao Shuai, Chi Zhang, Kaizhi Yang et al.

Unsupervised methods for reconstructing structures face significant challenges in capturing the geometric details with consistent structures among diverse shapes of the same category. To address this issue, we present a novel unsupervised structural reconstruction method, named DPF-Net, based on a new Deformable Primitive Field (DPF) representation, which allows for high-quality shape reconstruction using parameterized geometric primitives. We design a two-stage shape reconstruction pipeline which consists of a primitive generation module and a primitive deformation module to approximate the target shape of each part progressively. The primitive generation module estimates the explicit orientation, position, and size parameters of parameterized geometric primitives, while the primitive deformation module predicts a dense deformation field based on a parameterized primitive field to recover shape details. The strong shape prior encoded in parameterized geometric primitives enables our DPF-Net to extract high-level structures and recover fine-grained shape details consistently. The experimental results on three categories of objects in diverse shapes demonstrate the effectiveness and generalization ability of our DPF-Net on structural reconstruction and shape segmentation.

Binxin Yang, Xuejin Chen, Chaoqun Wang et al.

With recent advances in image-to-image translation tasks, remarkable progress has been witnessed in generating face images from sketches. However, existing methods frequently fail to generate images with details that are semantically and geometrically consistent with the input sketch, especially when various decoration strokes are drawn. To address this issue, we introduce a novel W-W+ encoder architecture to take advantage of the high expressive power of W+ space and semantic controllability of W space. We introduce an explicit intermediate representation for sketch semantic embedding. With a semantic feature matching loss for effective semantic supervision, our sketch embedding precisely conveys the semantics in the input sketches to the synthesized images. Moreover, a novel sketch semantic interpretation approach is designed to automatically extract semantics from vectorized sketches. We conduct extensive experiments on both synthesized sketches and hand-drawn sketches, and the results demonstrate the superiority of our method over existing approaches on both semantics-preserving and generalization ability.

Kai Cheng, Hao Chen, Wei Yin et al.

Multi-view depth estimation plays a critical role in reconstructing and understanding the 3D world. Recent learning-based methods have made significant progress in it. However, multi-view depth estimation is fundamentally a correspondence-based optimization problem, but previous learning-based methods mainly rely on predefined depth hypotheses to build correspondence as the cost volume and implicitly regularize it to fit depth prediction, deviating from the essence of iterative optimization based on stereo correspondence. Thus, they suffer unsatisfactory precision and generalization capability. In this paper, we are the first to explore more general image correlations to establish correspondences dynamically for depth estimation. We design a novel iterative multi-view depth estimation framework mimicking the optimization process, which consists of 1) a correlation volume construction module that models the pixel similarity between a reference image and source images as all-to-all correlations; 2) a flow-based depth initialization module that estimates the depth from the 2D optical flow; 3) a novel correlation-guided depth refinement module that reprojects points in different views to effectively fetch relevant correlations for further fusion and integrate the fused correlation for iterative depth update. Without predefined depth hypotheses, the fused correlations establish multi-view correspondence in an efficient way and guide the depth refinement heuristically. We conduct sufficient experiments on ScanNet, DeMoN, ETH3D, and 7Scenes to demonstrate the superiority of our method on multi-view depth estimation and its best generalization ability.

Xin Kang, Chaoqun Wang, Xuejin Chen

Semantic segmentation in complex scenes relies not only on object appearance but also on object location and the surrounding environment. Nonetheless, it is difficult to model long-range context in the format of pairwise point correlations due to the huge computational cost for large-scale point clouds. In this paper, we propose using regions as the intermediate representation of point clouds instead of fine-grained points or voxels to reduce the computational burden. We introduce a novel Region-Enhanced Feature Learning Network (REFL-Net) that leverages region correlations to enhance point feature learning. We design a region-based feature enhancement (RFE) module, which consists of a Semantic-Spatial Region Extraction stage and a Region Dependency Modeling stage. In the first stage, the input points are grouped into a set of regions based on their semantic and spatial proximity. In the second stage, we explore inter-region semantic and spatial relationships by employing a self-attention block on region features and then fuse point features with the region features to obtain more discriminative representations. Our proposed RFE module is plug-and-play and can be integrated with common semantic segmentation backbones. We conduct extensive experiments on ScanNetV2 and S3DIS datasets and evaluate our RFE module with different segmentation backbones. Our REFL-Net achieves 1.8% mIoU gain on ScanNetV2 and 1.7% mIoU gain on S3DIS with negligible computational cost compared with backbone models. Both quantitative and qualitative results show the powerful long-range context modeling ability and strong generalization ability of our REFL-Net.

Qihua Chen, Xuejin Chen, Hyun-Myung Woo et al.

Various real-world scientific applications involve the mathematical modeling of complex uncertain systems with numerous unknown parameters. Accurate parameter estimation is often practically infeasible in such systems, as the available training data may be insufficient and the cost of acquiring additional data may be high. In such cases, based on a Bayesian paradigm, we can design robust operators retaining the best overall performance across all possible models and design optimal experiments that can effectively reduce uncertainty to enhance the performance of such operators maximally. While objective-based uncertainty quantification (objective-UQ) based on MOCU (mean objective cost of uncertainty) provides an effective means for quantifying uncertainty in complex systems, the high computational cost of estimating MOCU has been a challenge in applying it to real-world scientific/engineering problems. In this work, we propose a novel scheme to reduce the computational cost for objective-UQ via MOCU based on a data-driven approach. We adopt a neural message-passing model for surrogate modeling, incorporating a novel axiomatic constraint loss that penalizes an increase in the estimated system uncertainty. As an illustrative example, we consider the optimal experimental design (OED) problem for uncertain Kuramoto models, where the goal is to predict the experiments that can most effectively enhance robust synchronization performance through uncertainty reduction. We show that our proposed approach can accelerate MOCU-based OED by four to five orders of magnitude, without any visible performance loss compared to the state-of-the-art. The proposed approach applies to general OED tasks, beyond the Kuramoto model.

Zhihua Cheng, Xuejin Chen

Sketching is an intuitive and effective way for content creation. While significant progress has been made for photorealistic image generation by using generative adversarial networks, it remains challenging to take a fine-grained control on synthetic content. The instance normalization layer, which is widely adopted in existing image translation networks, washes away details in the input sketch and leads to loss of precise control on the desired shape of the generated face images. In this paper, we comprehensively investigate the effect of instance normalization on generating photorealistic face images from hand-drawn sketches. We first introduce a visualization approach to analyze the feature embedding for sketches with a group of specific changes. Based on the visual analysis, we modify the instance normalization layers in the baseline image translation model. We elaborate a new set of hand-drawn sketches with 11 categories of specially designed changes and conduct extensive experimental analysis. The results and user studies demonstrate that our method markedly improve the quality of synthesized images and the conformance with user intention.

Jiaqi Xu, Yuwang Wang, Xuejin Chen

Domain generalization aims to address the domain shift between training and testing data. To learn the domain invariant representations, the model is usually trained on multiple domains. It has been found that the gradients of network weight relative to a specific task loss can characterize the task itself. In this work, with the assumption that the gradients of a specific domain samples under the classification task could also reflect the property of the domain, we propose a Shape Guided Gradient Voting (SGGV) method for domain generalization. Firstly, we introduce shape prior via extra inputs of the network to guide gradient descending towards a shape-biased direction for better generalization. Secondly, we propose a new gradient voting strategy to remove the outliers for robust optimization in the presence of shape guidance. To provide shape guidance, we add edge/sketch extracted from the training data as an explicit way, and also use texture augmented images as an implicit way. We conduct experiments on several popular domain generalization datasets in image classification task, and show that our shape guided gradient updating strategy brings significant improvement of the generalization.

Zhihong Zhang, Jie Zhao, Xiaojian Huang et al.

Multimodal reward models (MRMs) play a crucial role in aligning Multimodal Large Language Models (MLLMs) with human preferences. Training a good MRM requires high-quality multimodal preference data. However, existing preference datasets face three key challenges: lack of granularity in preference strength, textual style bias, and unreliable preference signals. Besides, existing open-source multimodal preference datasets suffer from substantial noise, yet there is a lack of effective and scalable curation methods to enhance their quality. To address these limitations, we propose \textbf{DT2IT-MRM}, which integrates a \textbf{D}ebiased preference construction pipeline, a novel reformulation of text-to-image (\textbf{T2I}) preference data, and an \textbf{I}terative \textbf{T}raining framework that curates existing multimodal preference datasets for \textbf{M}ultimodal \textbf{R}eward \textbf{M}odeling. Our experimental results show that DT2IT-MRM achieves new \textbf{state-of-the-art} overall performance on three major benchmarks: VL-RewardBench, Multimodal RewardBench, and MM-RLHF-RewardBench.

Qihua Chen, Xuejin Chen, Chenxuan Wang et al.

The current neuron reconstruction pipeline for electron microscopy (EM) data usually includes automatic image segmentation followed by extensive human expert proofreading. In this work, we aim to reduce human workload by predicting connectivity between over-segmented neuron pieces, taking both microscopy image and 3D morphology features into account, similar to human proofreading workflow. To this end, we first construct a dataset, named FlyTracing, that contains millions of pairwise connections of segments expanding the whole fly brain, which is three orders of magnitude larger than existing datasets for neuron segment connection. To learn sophisticated biological imaging features from the connectivity annotations, we propose a novel connectivity-aware contrastive learning method to generate dense volumetric EM image embedding. The learned embeddings can be easily incorporated with any point or voxel-based morphological representations for automatic neuron tracing. Extensive comparisons of different combination schemes of image and morphological representation in identifying split errors across the whole fly brain demonstrate the superiority of the proposed approach, especially for the locations that contain severe imaging artifacts, such as section missing and misalignment. The dataset and code are available at https://github.com/Levishery/Flywire-Neuron-Tracing.

Kai Cheng, Xiaoxiao Long, Kaizhi Yang et al.

The advent of 3D Gaussian Splatting (3DGS) has recently brought about a revolution in the field of neural rendering, facilitating high-quality renderings at real-time speed. However, 3DGS heavily depends on the initialized point cloud produced by Structure-from-Motion (SfM) techniques. When tackling with large-scale scenes that unavoidably contain texture-less surfaces, the SfM techniques always fail to produce enough points in these surfaces and cannot provide good initialization for 3DGS. As a result, 3DGS suffers from difficult optimization and low-quality renderings. In this paper, inspired by classical multi-view stereo (MVS) techniques, we propose GaussianPro, a novel method that applies a progressive propagation strategy to guide the densification of the 3D Gaussians. Compared to the simple split and clone strategies used in 3DGS, our method leverages the priors of the existing reconstructed geometries of the scene and patch matching techniques to produce new Gaussians with accurate positions and orientations. Experiments on both large-scale and small-scale scenes validate the effectiveness of our method, where our method significantly surpasses 3DGS on the Waymo dataset, exhibiting an improvement of 1.15dB in terms of PSNR.

Jiaqi Xu, Cuiling Lan, Xuejin Chen et al.

Human beings solve complex problems through critical thinking, where reasoning and evaluation are intertwined to converge toward correct solutions. However, most existing large language models (LLMs) decouple reasoning from verification: they either generate reasoning without explicit self-checking or rely on external verifiers to detect errors post hoc. The former lacks immediate feedback, while the latter increases system complexity and hinders synchronized learning. Motivated by human critical thinking, we propose Stepwise Think-Critique (STC), a unified framework that interleaves reasoning and self-critique at each step within a single model. STC is trained with a hybrid reinforcement learning objective combining reasoning rewards and critique-consistency rewards to jointly optimize reasoning quality and self-evaluation. Experiments on mathematical reasoning benchmarks show that STC demonstrates strong critic-thinking capabilities and produces more interpretable reasoning traces, representing a step toward LLMs with built-in critical thinking.

Zhihong Zhang, Xiaojian Huang, Jin Xu et al.

Multimodal reward models (MRMs) play a crucial role in the training, inference, and evaluation of Large Vision Language Models (LVLMs) by assessing response quality. However, existing benchmarks for evaluating MRMs in the video domain suffer from a limited number and diversity of questions, a lack of comprehensive evaluation dimensions, and inadequate evaluation of diverse types of MRMs. To address these gaps, we introduce VideoRewardBench, the first comprehensive benchmark covering four core aspects of video understanding: perception, knowledge, reasoning, and safety. Through our AI-assisted data pipeline, we curate a high-quality preference dataset of 1,563 annotated samples, including 1,482 unique videos and 1,559 distinct questions--15 times the number found in the most question-rich prior benchmark. Each sample is a triplet consisting of a video-text prompt, a chosen response, and a rejected response. We also conduct a comprehensive evaluation across 28 multimodal reward models spanning three categories: generative, discriminative, and semi-scalar. Results show that even the top-performing model GPT-4o achieves only 57.0% overall accuracy, and the state-of-the-art open-source model Qwen2.5-VL-72B reaches merely 53.3%. Our analysis further reveals three key insights: (i) MRMs trained with reinforcement learning (RL) do not necessarily exhibit stronger cross-modal generalization than those trained without RL; (ii) except for discriminative MRMs, other types of MRMs across varying model capacities can benefit from inference-time scaling; and (iii) variations in input video frame count have different effects on different types of MRMs. We believe VideoRewardBench offers a challenging and valuable benchmark for advancing the evaluation and development of MRMs in the video domain.

Siqi Chen, Guoqing Zhang, Jiahao Lai et al.

Advancements in 3D vision have increased the impact of blood vessel modeling on medical applications. However, accurately representing the complex geometry and topology of blood vessels remains a challenge due to their intricate branching patterns, curvatures, and irregular shapes. In this study, we propose a hierarchical part-based frame work for 3D vessel generation that separates the global binary tree-like topology from local geometric details. Our approach proceeds in three stages: (1) key graph generation to model the overall hierarchical struc ture, (2) vessel segment generation conditioned on geometric properties, and (3) hierarchical vessel assembly by integrating the local segments according to the global key graph. We validate our framework on real world datasets, demonstrating superior performance over existing methods in modeling complex vascular networks. This work marks the first successful application of a part-based generative approach for 3D vessel modeling, setting a new benchmark for vascular data generation. The code is available at: https://github.com/CybercatChen/PartVessel.git.

Xiaotian Chen, Yuwang Wang, Xuejin Chen et al.

Human can infer the 3D geometry of a scene from a sketch instead of a realistic image, which indicates that the spatial structure plays a fundamental role in understanding the depth of scenes. We are the first to explore the learning of a depth-specific structural representation, which captures the essential feature for depth estimation and ignores irrelevant style information. Our S2R-DepthNet (Synthetic to Real DepthNet) can be well generalized to unseen real-world data directly even though it is only trained on synthetic data. S2R-DepthNet consists of: a) a Structure Extraction (STE) module which extracts a domaininvariant structural representation from an image by disentangling the image into domain-invariant structure and domain-specific style components, b) a Depth-specific Attention (DSA) module, which learns task-specific knowledge to suppress depth-irrelevant structures for better depth estimation and generalization, and c) a depth prediction module (DP) to predict depth from the depth-specific representation. Without access of any real-world images, our method even outperforms the state-of-the-art unsupervised domain adaptation methods which use real-world images of the target domain for training. In addition, when using a small amount of labeled real-world data, we achieve the state-ofthe-art performance under the semi-supervised setting. The code and trained models are available at https://github.com/microsoft/S2R-DepthNet.

Xiaotian Chen, Xuejin Chen, Zheng-Jun Zha

Monocular depth estimation is an essential task for scene understanding. The underlying structure of objects and stuff in a complex scene is critical to recovering accurate and visually-pleasing depth maps. Global structure conveys scene layouts, while local structure reflects shape details. Recently developed approaches based on convolutional neural networks (CNNs) significantly improve the performance of depth estimation. However, few of them take into account multi-scale structures in complex scenes. In this paper, we propose a Structure-Aware Residual Pyramid Network (SARPN) to exploit multi-scale structures for accurate depth prediction. We propose a Residual Pyramid Decoder (RPD) which expresses global scene structure in upper levels to represent layouts, and local structure in lower levels to present shape details. At each level, we propose Residual Refinement Modules (RRM) that predict residual maps to progressively add finer structures on the coarser structure predicted at the upper level. In order to fully exploit multi-scale image features, an Adaptive Dense Feature Fusion (ADFF) module, which adaptively fuses effective features from all scales for inferring structures of each scale, is introduced. Experiment results on the challenging NYU-Depth v2 dataset demonstrate that our proposed approach achieves state-of-the-art performance in both qualitative and quantitative evaluation. The code is available at https://github.com/Xt-Chen/SARPN.

Jiaqi Xu, Cuiling Lan, Wenxuan Xie et al.

Video-Language Models (VLMs), powered by the advancements in Large Language Models (LLMs), are charting new frontiers in video understanding. A pivotal challenge is the development of an efficient method to encapsulate video content into a set of representative tokens to align with LLMs. In this work, we introduce Slot-VLM, a novel framework designed to generate semantically decomposed video tokens, in terms of object-wise and event-wise visual representations, to facilitate LLM inference. Particularly, we design a SlowFast Slots module, i.e., SF-Slots, that adaptively aggregates the dense video tokens from the CLIP vision encoder to a set of representative slots. In order to take into account both the spatial object details and the varied temporal dynamics, SF-Slots is built with a dual-branch structure. The Slow-Slots branch focuses on extracting object-centric slots from features at high spatial resolution but low (slow) frame sample rate, emphasizing detailed object information. Conversely, Fast-Slots branch is engineered to learn event-centric slots from high temporal sample rate but low spatial resolution features. These complementary slots are combined to form the vision context, serving as the input to the LLM for efficient question answering. Our experimental results demonstrate the effectiveness of our Slot-VLM, which achieves the state-of-the-art performance on video question-answering.

Jiaqi Xu, Cuiling Lan, Wenxuan Xie et al.

The remarkable natural language understanding, reasoning, and generation capabilities of large language models (LLMs) have made them attractive for application to video understanding, utilizing video tokens as contextual input. However, employing LLMs for long video understanding presents significant challenges. The extensive number of video tokens leads to considerable computational costs for LLMs while using aggregated tokens results in loss of vision details. Moreover, the presence of abundant question-irrelevant tokens introduces noise to the video reasoning process. To address these issues, we introduce a simple yet effective learnable retrieval-based video-language model (R-VLM) for efficient long video understanding. Specifically, given a question (query) and a long video, our model identifies and selects the most relevant K video chunks and uses their associated visual tokens to serve as context for the LLM inference. This effectively reduces the number of video tokens, eliminates noise interference, and enhances system performance. We achieve this by incorporating a learnable lightweight MLP block to facilitate the efficient retrieval of question-relevant chunks, through the end-to-end training of our video-language model with a proposed soft matching loss. Our experimental results on multiple zero-shot video question answering datasets validate the effectiveness of our framework for comprehending long videos.

Yanzhe Lyu, Kai Cheng, Xin Kang et al.

Recently, 3D Gaussian Splatting (3D-GS) has prevailed in novel view synthesis, achieving high fidelity and efficiency. However, it often struggles to capture rich details and complete geometry. Our analysis reveals that the 3D-GS densification operation lacks adaptiveness and faces a dilemma between geometry coverage and detail recovery. To address this, we introduce a novel densification operation, residual split, which adds a downscaled Gaussian as a residual. Our approach is capable of adaptively retrieving details and complementing missing geometry. To further support this method, we propose a pipeline named ResGS. Specifically, we integrate a Gaussian image pyramid for progressive supervision and implement a selection scheme that prioritizes the densification of coarse Gaussians over time. Extensive experiments demonstrate that our method achieves SOTA rendering quality. Consistent performance improvements can be achieved by applying our residual split on various 3D-GS variants, underscoring its versatility and potential for broader application in 3D-GS-based applications.

Xin Kang, Zihan Zheng, Lei Chu et al.

We present LTM3D, a Latent Token space Modeling framework for conditional 3D shape generation that integrates the strengths of diffusion and auto-regressive (AR) models. While diffusion-based methods effectively model continuous latent spaces and AR models excel at capturing inter-token dependencies, combining these paradigms for 3D shape generation remains a challenge. To address this, LTM3D features a Conditional Distribution Modeling backbone, leveraging a masked autoencoder and a diffusion model to enhance token dependency learning. Additionally, we introduce Prefix Learning, which aligns condition tokens with shape latent tokens during generation, improving flexibility across modalities. We further propose a Latent Token Reconstruction module with Reconstruction-Guided Sampling to reduce uncertainty and enhance structural fidelity in generated shapes. Our approach operates in token space, enabling support for multiple 3D representations, including signed distance fields, point clouds, meshes, and 3D Gaussian Splatting. Extensive experiments on image- and text-conditioned shape generation tasks demonstrate that LTM3D outperforms existing methods in prompt fidelity and structural accuracy while offering a generalizable framework for multi-modal, multi-representation 3D generation.

Yue Xu, Kaizhi Yang, Jiebo Luo et al.

3D visual grounding is an emerging research area dedicated to making connections between the 3D physical world and natural language, which is crucial for achieving embodied intelligence. In this paper, we propose DASANet, a Dual Attribute-Spatial relation Alignment Network that separately models and aligns object attributes and spatial relation features between language and 3D vision modalities. We decompose both the language and 3D point cloud input into two separate parts and design a dual-branch attention module to separately model the decomposed inputs while preserving global context in attribute-spatial feature fusion by cross attentions. Our DASANet achieves the highest grounding accuracy 65.1% on the Nr3D dataset, 1.3% higher than the best competitor. Besides, the visualization of the two branches proves that our method is efficient and highly interpretable.

Chaoqun Wang, Shaobo Min, Xuejin Chen et al.

Generalized Zero-Shot Learning (GZSL) aims to recognize new categories with auxiliary semantic information,e.g., category attributes. In this paper, we handle the critical issue of domain shift problem, i.e., confusion between seen and unseen categories, by progressively improving cross-domain transferability and category discriminability of visual representations. Our approach, named Dual Progressive Prototype Network (DPPN), constructs two types of prototypes that record prototypical visual patterns for attributes and categories, respectively. With attribute prototypes, DPPN alternately searches attribute-related local regions and updates corresponding attribute prototypes to progressively explore accurate attribute-region correspondence. This enables DPPN to produce visual representations with accurate attribute localization ability, which benefits the semantic-visual alignment and representation transferability. Besides, along with progressive attribute localization, DPPN further projects category prototypes into multiple spaces to progressively repel visual representations from different categories, which boosts category discriminability. Both attribute and category prototypes are collaboratively learned in a unified framework, which makes visual representations of DPPN transferable and distinctive. Experiments on four benchmarks prove that DPPN effectively alleviates the domain shift problem in GZSL.

Zhili Li, Xuejin Chen, Jie Zhao et al.

Mitochondria segmentation in electron microscopy images is essential in neuroscience. However, due to the image degradation during the imaging process, the large variety of mitochondrial structures, as well as the presence of noise, artifacts and other sub-cellular structures, mitochondria segmentation is very challenging. In this paper, we propose a novel and effective contrastive learning framework to learn a better feature representation from hard examples to improve segmentation. Specifically, we adopt a point sampling strategy to pick out representative pixels from hard examples in the training phase. Based on these sampled pixels, we introduce a pixel-wise label-based contrastive loss which consists of a similarity loss term and a consistency loss term. The similarity term can increase the similarity of pixels from the same class and the separability of pixels from different classes in feature space, while the consistency term is able to enhance the sensitivity of the 3D model to changes in image content from frame to frame. We demonstrate the effectiveness of our method on MitoEM dataset as well as FIB-SEM dataset and show better or on par with state-of-the-art results.

Hao Wang, Zheng-Jun Zha, Liang Li et al.

We study the problem of localizing audio-visual events that are both audible and visible in a video. Existing works focus on encoding and aligning audio and visual features at the segment level while neglecting informative correlation between segments of the two modalities and between multi-scale event proposals. We propose a novel MultiModulation Network (M2N) to learn the above correlation and leverage it as semantic guidance to modulate the related auditory, visual, and fused features. In particular, during feature encoding, we propose cross-modal normalization and intra-modal normalization. The former modulates the features of two modalities by establishing and exploiting the cross-modal relationship. The latter modulates the features of a single modality with the event-relevant semantic guidance of the same modality. In the fusion stage,we propose a multi-scale proposal modulating module and a multi-alignment segment modulating module to introduce multi-scale event proposals and enable dense matching between cross-modal segments. With the auditory, visual, and fused features modulated by the correlation information regarding audio-visual events, M2N performs accurate event localization. Extensive experiments conducted on the AVE dataset demonstrate that our proposed method outperforms the state of the art in both supervised event localization and cross-modality localization.

Kaizhi Yang, Xuejin Chen

Representing complex 3D objects as simple geometric primitives, known as shape abstraction, is important for geometric modeling, structural analysis, and shape synthesis. In this paper, we propose an unsupervised shape abstraction method to map a point cloud into a compact cuboid representation. We jointly predict cuboid allocation as part segmentation and cuboid shapes and enforce the consistency between the segmentation and shape abstraction for self-learning. For the cuboid abstraction task, we transform the input point cloud into a set of parametric cuboids using a variational auto-encoder network. The segmentation network allocates each point into a cuboid considering the point-cuboid affinity. Without manual annotations of parts in point clouds, we design four novel losses to jointly supervise the two branches in terms of geometric similarity and cuboid compactness. We evaluate our method on multiple shape collections and demonstrate its superiority over existing shape abstraction methods. Moreover, based on our network architecture and learned representations, our approach supports various applications including structured shape generation, shape interpolation, and structural shape clustering.

Chaoqun Wang, Xuejin Chen, Shaobo Min et al.

Generalized Zero-Shot Learning (GZSL) targets recognizing new categories by learning transferable image representations. Existing methods find that, by aligning image representations with corresponding semantic labels, the semantic-aligned representations can be transferred to unseen categories. However, supervised by only seen category labels, the learned semantic knowledge is highly task-specific, which makes image representations biased towards seen categories. In this paper, we propose a novel Dual-Contrastive Embedding Network (DCEN) that simultaneously learns task-specific and task-independent knowledge via semantic alignment and instance discrimination. First, DCEN leverages task labels to cluster representations of the same semantic category by cross-modal contrastive learning and exploring semantic-visual complementarity. Besides task-specific knowledge, DCEN then introduces task-independent knowledge by attracting representations of different views of the same image and repelling representations of different images. Compared to high-level seen category supervision, this instance discrimination supervision encourages DCEN to capture low-level visual knowledge, which is less biased toward seen categories and alleviates the representation bias. Consequently, the task-specific and task-independent knowledge jointly make for transferable representations of DCEN, which obtains averaged 4.1% improvement on four public benchmarks.

Songfang Han, Jiayuan Gu, Kaichun Mo et al.

Single-image 3D shape reconstruction is an important and long-standing problem in computer vision. A plethora of existing works is constantly pushing the state-of-the-art performance in the deep learning era. However, there remains a much more difficult and under-explored issue on how to generalize the learned skills over unseen object categories that have very different shape geometry distributions. In this paper, we bring in the concept of compositional generalizability and propose a novel framework that could better generalize to these unseen categories. We factorize the 3D shape reconstruction problem into proper sub-problems, each of which is tackled by a carefully designed neural sub-module with generalizability concerns. The intuition behind our formulation is that object parts (slates and cylindrical parts), their relationships (adjacency and translation symmetry), and shape substructures (T-junctions and a symmetric group of parts) are mostly shared across object categories, even though object geometries may look very different (e.g. chairs and cabinets). Experiments on PartNet show that we achieve superior performance than state-of-the-art. This validates our problem factorization and network designs.

Yuhang Li, Xuejin Chen, Binxin Yang et al.

In this paper, we explore the task of generating photo-realistic face images from hand-drawn sketches. Existing image-to-image translation methods require a large-scale dataset of paired sketches and images for supervision. They typically utilize synthesized edge maps of face images as training data. However, these synthesized edge maps strictly align with the edges of the corresponding face images, which limit their generalization ability to real hand-drawn sketches with vast stroke diversity. To address this problem, we propose DeepFacePencil, an effective tool that is able to generate photo-realistic face images from hand-drawn sketches, based on a novel dual generator image translation network during training. A novel spatial attention pooling (SAP) is designed to adaptively handle stroke distortions which are spatially varying to support various stroke styles and different levels of details. We conduct extensive experiments and the results demonstrate the superiority of our model over existing methods on both image quality and model generalization to hand-drawn sketches.

Yuhang Li, Xuejin Chen, Feng Wu et al.

In this paper, we explore the task of generating photo-realistic face images from lines. Previous methods based on conditional generative adversarial networks (cGANs) have shown their power to generate visually plausible images when a conditional image and an output image share well-aligned structures. However, these models fail to synthesize face images with a whole set of well-defined structures, e.g. eyes, noses, mouths, etc., especially when the conditional line map lacks one or several parts. To address this problem, we propose a conditional self-attention generative adversarial network (CSAGAN). We introduce a conditional self-attention mechanism to cGANs to capture long-range dependencies between different regions in faces. We also build a multi-scale discriminator. The large-scale discriminator enforces the completeness of global structures and the small-scale discriminator encourages fine details, thereby enhancing the realism of generated face images. We evaluate the proposed model on the CelebA-HD dataset by two perceptual user studies and three quantitative metrics. The experiment results demonstrate that our method generates high-quality facial images while preserving facial structures. Our results outperform state-of-the-art methods both quantitatively and qualitatively.

Shaobo Min, Xuejin Chen, Zheng-Jun Zha et al.

\begin{abstract} Learning-based methods suffer from a deficiency of clean annotations, especially in biomedical segmentation. Although many semi-supervised methods have been proposed to provide extra training data, automatically generated labels are usually too noisy to retrain models effectively. In this paper, we propose a Two-Stream Mutual Attention Network (TSMAN) that weakens the influence of back-propagated gradients caused by incorrect labels, thereby rendering the network robust to unclean data. The proposed TSMAN consists of two sub-networks that are connected by three types of attention models in different layers. The target of each attention model is to indicate potentially incorrect gradients in a certain layer for both sub-networks by analyzing their inferred features using the same input. In order to achieve this purpose, the attention models are designed based on the propagation analysis of noisy gradients at different layers. This allows the attention models to effectively discover incorrect labels and weaken their influence during the parameter updating process. By exchanging multi-level features within the two-stream architecture, the effects of noisy labels in each sub-network are reduced by decreasing the updating gradients. Furthermore, a hierarchical distillation is developed to provide more reliable pseudo labels for unlabelded data, which further boosts the performance of our retrained TSMAN. The experiments using both the HVSMR 2016 and BRATS 2015 benchmarks demonstrate that our semi-supervised learning framework surpasses the state-of-the-art fully-supervised results.