Zhe Huang

Chengwen Liu, Xiaomin Yu, Zhuoyue Chang et al.

In real-world video question answering scenarios, videos often provide only localized visual cues, while verifiable answers are distributed across the open web; models therefore need to jointly perform cross-frame clue extraction, iterative retrieval, and multi-hop reasoning-based verification. To bridge this gap, we construct the first video deep research benchmark, VideoDR. VideoDR centers on video-conditioned open-domain video question answering, requiring cross-frame visual anchor extraction, interactive web retrieval, and multi-hop reasoning over joint video-web evidence; through rigorous human annotation and quality control, we obtain high-quality video deep research samples spanning six semantic domains. We evaluate multiple closed-source and open-source multimodal large language models under both the Workflow and Agentic paradigms, and the results show that Agentic is not consistently superior to Workflow: its gains depend on a model's ability to maintain the initial video anchors over long retrieval chains. Further analysis indicates that goal drift and long-horizon consistency are the core bottlenecks. In sum, VideoDR provides a systematic benchmark for studying video agents in open-web settings and reveals the key challenges for next-generation video deep research agents.

Adam Villaflor, Zhe Huang, Swapnil Pande et al.

Impressive results in natural language processing (NLP) based on the Transformer neural network architecture have inspired researchers to explore viewing offline reinforcement learning (RL) as a generic sequence modeling problem. Recent works based on this paradigm have achieved state-of-the-art results in several of the mostly deterministic offline Atari and D4RL benchmarks. However, because these methods jointly model the states and actions as a single sequencing problem, they struggle to disentangle the effects of the policy and world dynamics on the return. Thus, in adversarial or stochastic environments, these methods lead to overly optimistic behavior that can be dangerous in safety-critical systems like autonomous driving. In this work, we propose a method that addresses this optimism bias by explicitly disentangling the policy and world models, which allows us at test time to search for policies that are robust to multiple possible futures in the environment. We demonstrate our method's superior performance on a variety of autonomous driving tasks in simulation.

Shuijing Liu, Peixin Chang, Zhe Huang et al.

We study the problem of safe and intention-aware robot navigation in dense and interactive crowds. Most previous reinforcement learning (RL) based methods fail to consider different types of interactions among all agents or ignore the intentions of people, which results in performance degradation. To learn a safe and efficient robot policy, we propose a novel recurrent graph neural network with attention mechanisms to capture heterogeneous interactions among agents through space and time. To encourage longsighted robot behaviors, we infer the intentions of dynamic agents by predicting their future trajectories for several timesteps. The predictions are incorporated into a model-free RL framework to prevent the robot from intruding into the intended paths of other agents. We demonstrate that our method enables the robot to achieve good navigation performance and non-invasiveness in challenging crowd navigation scenarios. We successfully transfer the policy learned in simulation to a real-world TurtleBot 2i. Our code and videos are available at https://sites.google.com/view/intention-aware-crowdnav/home.

Zhe Huang, Shuo Wang, Yongcai Wang et al.

Collaborative autonomous driving with multiple vehicles usually requires the data fusion from multiple modalities. To ensure effective fusion, the data from each individual modality shall maintain a reasonably high quality. However, in collaborative perception, the quality of object detection based on a modality is highly sensitive to the relative pose errors among the agents. It leads to feature misalignment and significantly reduces collaborative performance. To address this issue, we propose RoCo, a novel unsupervised framework to conduct iterative object matching and agent pose adjustment. To the best of our knowledge, our work is the first to model the pose correction problem in collaborative perception as an object matching task, which reliably associates common objects detected by different agents. On top of this, we propose a graph optimization process to adjust the agent poses by minimizing the alignment errors of the associated objects, and the object matching is re-done based on the adjusted agent poses. This process is carried out iteratively until convergence. Experimental study on both simulated and real-world datasets demonstrates that the proposed framework RoCo consistently outperforms existing relevant methods in terms of the collaborative object detection performance, and exhibits highly desired robustness when the pose information of agents is with high-level noise. Ablation studies are also provided to show the impact of its key parameters and components. The code is released at https://github.com/HuangZhe885/RoCo.

Yihao Ding, Zhe Huang, Runlin Wang et al.

We propose V-Doc, a question-answering tool using document images and PDF, mainly for researchers and general non-deep learning experts looking to generate, process, and understand the document visual question answering tasks. The V-Doc supports generating and using both extractive and abstractive question-answer pairs using documents images. The extractive QA selects a subset of tokens or phrases from the document contents to predict the answers, while the abstractive QA recognises the language in the content and generates the answer based on the trained model. Both aspects are crucial to understanding the documents, especially in an image format. We include a detailed scenario of question generation for the abstractive QA task. V-Doc supports a wide range of datasets and models, and is highly extensible through a declarative, framework-agnostic platform.

Shuo Liu, Zhe Huang, Calin A. Belta

Obstacle avoidance of polytopic obstacles by polytopic robots is a challenging problem in optimization-based control and trajectory planning. Many existing methods rely on smooth geometric approximations, such as hyperspheres or ellipsoids, which allow differentiable distance expressions but distort the true geometry and restrict the feasible set. Other approaches integrate exact polytope distances into nonlinear model predictive control (MPC), resulting in nonconvex programs that limit real-time performance. In this paper, we construct linear discrete-time control barrier function (DCBF) constraints by deriving supporting hyperplanes from exact closest-point computations between convex polytopes. We then propose a novel iterative convex MPC-DCBF framework, where local linearization of system dynamics and robot geometry ensures convexity of the finite-horizon optimization at each iteration. The resulting formulation reduces computational complexity and enables fast online implementation for safety-critical control and trajectory planning of general nonlinear dynamics. The framework extends to multi-robot and three-dimensional environments. Numerical experiments demonstrate collision-free navigation in cluttered maze scenarios with millisecond-level solve times.

Hongyu Sun, Yongcai Wang, Xudong Cai et al.

One fundamental limitation to the research of bird strike prevention is the lack of a large-scale dataset taken directly from real-world airports. Existing relevant datasets are either small in size or not dedicated for this purpose. To advance the research and practical solutions for bird strike prevention, in this paper, we present a large-scale challenging dataset AirBirds that consists of 118,312 time-series images, where a total of 409,967 bounding boxes of flying birds are manually, carefully annotated. The average size of all annotated instances is smaller than 10 pixels in 1920x1080 images. Images in the dataset are captured over 4 seasons of a whole year by a network of cameras deployed at a real-world airport, covering diverse bird species, lighting conditions and 13 meteorological scenarios. To the best of our knowledge, it is the first large-scale image dataset that directly collects flying birds in real-world airports for bird strike prevention. This dataset is publicly available at https://airbirdsdata.github.io/.

Zhe Huang, Mary-Joy Sidhom, Benjamin S. Wessler et al.

Semi-supervised learning (SSL) promises improved accuracy compared to training classifiers on small labeled datasets by also training on many unlabeled images. In real applications like medical imaging, unlabeled data will be collected for expediency and thus uncurated: possibly different from the labeled set in classes or features. Unfortunately, modern deep SSL often makes accuracy worse when given uncurated unlabeled data. Recent complex remedies try to detect out-of-distribution unlabeled images and then discard or downweight them. Instead, we introduce Fix-A-Step, a simpler procedure that views all uncurated unlabeled images as potentially helpful. Our first insight is that even uncurated images can yield useful augmentations of labeled data. Second, we modify gradient descent updates to prevent optimizing a multi-task SSL loss from hurting labeled-set accuracy. Fix-A-Step can repair many common deep SSL methods, improving accuracy on CIFAR benchmarks across all tested methods and levels of artificial class mismatch. On a new medical SSL benchmark called Heart2Heart, Fix-A-Step can learn from 353,500 truly uncurated ultrasound images to deliver gains that generalize across hospitals.

Yan Zhang, Long Ma, Yuxin Feng et al.

Learning-based real image dehazing methods have achieved notable progress, yet they still face adaptation challenges in diverse real haze scenes. These challenges mainly stem from the lack of effective unsupervised mechanisms for unlabeled data and the heavy cost of full model fine-tuning. To address these challenges, we propose the haze-to-clear text-directed loss that leverages CLIP's cross-modal capabilities to reformulate real image dehazing as a semantic alignment problem in latent space, thereby providing explicit unsupervised cross-modal guidance in the absence of reference images. Furthermore, we introduce the Bilevel Layer-positioning LoRA (BiLaLoRA) strategy, which learns both the LoRA parameters and automatically search the injection layers, enabling targeted adaptation of critical network layers. Extensive experiments demonstrate our superiority against state-of-the-art methods on multiple real-world dehazing benchmarks. The code is publicly available at https://github.com/YanZhang-zy/BiLaLoRA.

Zhe Huang, Hao Wen, Aiming Hao et al.

Multimodal Large Language Models (MLLMs) have made remarkable progress in video understanding. However, they suffer from a critical vulnerability: an over-reliance on language priors, which can lead to visual ungrounded hallucinations, especially when processing counterfactual videos that defy common sense. This limitation, stemming from the intrinsic data imbalance between text and video, is challenging to address due to the substantial cost of collecting and annotating counterfactual data. To address this, we introduce DualityForge, a novel counterfactual data synthesis framework that employs controllable, diffusion-based video editing to transform real-world videos into counterfactual scenarios. By embedding structured contextual information into the video editing and QA generation processes, the framework automatically produces high-quality QA pairs together with original-edited video pairs for contrastive training. Based on this, we build DualityVidQA, a large-scale video dataset designed to reduce MLLM hallucinations. In addition, to fully exploit the contrastive nature of our paired data, we propose Duality-Normalized Advantage Training (DNA-Train), a two-stage SFT-RL training regime where the RL phase applies pair-wise $\ell_1$ advantage normalization, thereby enabling a more stable and efficient policy optimization. Experiments on DualityVidQA-Test demonstrate that our method substantially reduces model hallucinations on counterfactual videos, yielding a relative improvement of 24.0% over the Qwen2.5-VL-7B baseline. Moreover, our approach achieves significant gains across both hallucination and general-purpose benchmarks, indicating strong generalization capability. We will open-source our dataset and code.

Soyeon Caren Han, Yihao Ding, Siwen Luo et al.

Document understanding and information extraction include different tasks to understand a document and extract valuable information automatically. Recently, there has been a rising demand for developing document understanding among different domains, including business, law, and medicine, to boost the efficiency of work that is associated with a large number of documents. This workshop aims to bring together researchers and industry developers in the field of document intelligence and understanding diverse document types to boost automatic document processing and understanding techniques. We also released a data challenge on the recently introduced document-level VQA dataset, PDFVQA. The PDFVQA challenge examines the structural and contextual understandings of proposed models on the natural full document level of multiple consecutive document pages by including questions with a sequence of answers extracted from multi-pages of the full document. This task helps to boost the document understanding step from the single-page level to the full document level understanding.

Xiaowei Yu, Zhe Huang, Minheng Chen et al.

We investigate the impact of entropy change in deep learning systems by noise injection at different levels, including the embedding space and the image. The series of models that employ our methodology are collectively known as Noisy Neural Networks (NoisyNN), with examples such as NoisyViT and NoisyCNN. Noise is conventionally viewed as a harmful perturbation in various deep learning architectures, such as convolutional neural networks (CNNs) and vision transformers (ViTs), as well as different learning tasks like image classification and transfer learning. However, this work shows noise can be an effective way to change the entropy of the learning system. We demonstrate that specific noise can boost the performance of various deep models under certain conditions. We theoretically prove the enhancement gained from positive noise by reducing the task complexity defined by information entropy and experimentally show the significant performance gain in large image datasets, such as the ImageNet. Herein, we use the information entropy to define the complexity of the task. We categorize the noise into two types, positive noise (PN) and harmful noise (HN), based on whether the noise can help reduce the task complexity. Extensive experiments of CNNs and ViTs have shown performance improvements by proactively injecting positive noise, where we achieved an unprecedented top 1 accuracy of 95$\%$ on ImageNet. Both theoretical analysis and empirical evidence have confirmed that the presence of positive noise, can benefit the learning process, while the traditionally perceived harmful noise indeed impairs deep learning models. The different roles of noise offer new explanations for deep models on specific tasks and provide a new paradigm for improving model performance. Moreover, it reminds us that we can influence the performance of learning systems via information entropy change.

Zhe Huang, Ruijie Jiang, Shuchin Aeron et al.

In typical medical image classification problems, labeled data is scarce while unlabeled data is more available. Semi-supervised learning and self-supervised learning are two different research directions that can improve accuracy by learning from extra unlabeled data. Recent methods from both directions have reported significant gains on traditional benchmarks. Yet past benchmarks do not focus on medical tasks and rarely compare self- and semi- methods together on an equal footing. Furthermore, past benchmarks often handle hyperparameter tuning suboptimally. First, they may not tune hyperparameters at all, leading to underfitting. Second, when tuning does occur, it often unrealistically uses a labeled validation set that is much larger than the training set. Therefore currently published rankings might not always corroborate with their practical utility This study contributes a systematic evaluation of self- and semi- methods with a unified experimental protocol intended to guide a practitioner with scarce overall labeled data and a limited compute budget. We answer two key questions: Can hyperparameter tuning be effective with realistic-sized validation sets? If so, when all methods are tuned well, which self- or semi-supervised methods achieve the best accuracy? Our study compares 13 representative semi- and self-supervised methods to strong labeled-set-only baselines on 4 medical datasets. From 20000+ GPU hours of computation, we provide valuable best practices to resource-constrained practitioners: hyperparameter tuning is effective, and the semi-supervised method known as MixMatch delivers the most reliable gains across 4 datasets.

Peihua Mai, Ran Yan, Zhe Huang et al.

Large Language Models (LLMs) excel in natural language understanding by capturing hidden semantics in vector space. This process enriches the value of text embeddings for various downstream tasks, thereby fostering the Embedding-as-a-Service (EaaS) business model. However, the risk of privacy leakage due to direct text transmission to servers remains a critical concern. To address this, we introduce Split-N-Denoise (SnD), an private inference framework that splits the model to execute the token embedding layer on the client side at minimal computational cost. This allows the client to introduce noise prior to transmitting the embeddings to the server, and subsequently receive and denoise the perturbed output embeddings for downstream tasks. Our approach is designed for the inference stage of LLMs and requires no modifications to the model parameters. Extensive experiments demonstrate SnD's effectiveness in optimizing the privacy-utility tradeoff across various LLM architectures and diverse downstream tasks. The results reveal an improvement in performance under the same privacy budget compared to the baselines by over 10\% on average, offering clients a privacy-preserving solution for local privacy protection.

Shuo Wang, Yongcai Wang, Zhimin Xu et al.

For interacting with mobile objects in unfamiliar environments, simultaneously locating, mapping, and tracking the 3D poses of multiple objects are crucially required. This paper proposes a Tracklet Graph and Query Graph-based framework, i.e., GSLAMOT, to address this challenge. GSLAMOT utilizes camera and LiDAR multimodal information as inputs and divides the representation of the dynamic scene into a semantic map for representing the static environment, a trajectory of the ego-agent, and an online maintained Tracklet Graph (TG) for tracking and predicting the 3D poses of the detected mobile objects. A Query Graph (QG) is constructed in each frame by object detection to query and update TG. For accurate object association, a Multi-criteria Star Graph Association (MSGA) method is proposed to find matched objects between the detections in QG and the predicted tracklets in TG. Then, an Object-centric Graph Optimization (OGO) method is proposed to simultaneously optimize the TG, the semantic map, and the agent trajectory. It triangulates the detected objects into the map to enrich the map's semantic information. We address the efficiency issues to handle the three tightly coupled tasks in parallel. Experiments are conducted on KITTI, Waymo, and an emulated Traffic Congestion dataset that highlights challenging scenarios. Experiments show that GSLAMOT enables accurate crowded object tracking while conducting SLAM accurately in challenging scenarios, demonstrating more excellent performances than the state-of-the-art methods. The code and dataset are at https://gslamot.github.io.

Zhe Huang, Peng Wang, Yan Zheng et al.

Product bundling boosts e-commerce revenue by recommending complementary item combinations. However, existing methods face two critical challenges: (1) collaborative filtering approaches struggle with cold-start items owing to dependency on historical interactions, and (2) LLMs lack inherent capability to model interactive graph directly. To bridge this gap, we propose a dual-enhancement method that integrates interactive graph learning and LLM-based semantic understanding for product bundling. Our method introduces a graph-to-text paradigm, which leverages a Dynamic Concept Binding Mechanism (DCBM) to translate graph structures into natural language prompts. The DCBM plays a critical role in aligning domain-specific entities with LLM tokenization, enabling effective comprehension of combinatorial constraints. Experiments on three benchmarks (POG, POG_dense, Steam) demonstrate 6.3%-26.5% improvements over state-of-the-art baselines.

Neeloy Chakraborty, Yixiao Fang, Andre Schreiber et al.

Teleoperation is an important technology to enable supervisors to control agricultural robots remotely. However, environmental factors in dense crop rows and limitations in network infrastructure hinder the reliability of data streamed to teleoperators. These issues result in delayed and variable frame rate video feeds that often deviate significantly from the robot's actual viewpoint. We propose a modular learning-based vision pipeline to generate delay-compensated images in real-time for supervisors. Our extensive offline evaluations demonstrate that our method generates more accurate images compared to state-of-the-art approaches in our setting. Additionally, ours is one of the few works to evaluate a delay-compensation method in outdoor field environments with complex terrain on data from a real robot in real-time. Resulting videos and code are provided at https://sites.google.com/illinois.edu/comp-teleop.

Xudong Cai, Yongcai Wang, Zhe Huang et al.

The availability of city-scale Lidar maps enables the potential of city-scale place recognition using mobile cameras. However, the city-scale Lidar maps generally need to be compressed for storage efficiency, which increases the difficulty of direct visual place recognition in compressed Lidar maps. This paper proposes VOLoc, an accurate and efficient visual place recognition method that exploits geometric similarity to directly query the compressed Lidar map via the real-time captured image sequence. In the offline phase, VOLoc compresses the Lidar maps using a \emph{Geometry-Preserving Compressor} (GPC), in which the compression is reversible, a crucial requirement for the downstream 6DoF pose estimation. In the online phase, VOLoc proposes an online Geometric Recovery Module (GRM), which is composed of online Visual Odometry (VO) and a point cloud optimization module, such that the local scene structure around the camera is online recovered to build the \emph{Querying Point Cloud} (QPC). Then the QPC is compressed by the same GPC, and is aggregated into a global descriptor by an attention-based aggregation module, to query the compressed Lidar map in the vector space. A transfer learning mechanism is also proposed to improve the accuracy and the generality of the aggregation network. Extensive evaluations show that VOLoc provides localization accuracy even better than the Lidar-to-Lidar place recognition, setting up a new record for utilizing the compressed Lidar map by low-end mobile cameras. The code are publicly available at https://github.com/Master-cai/VOLoc.

Zhe Huang, Shuo Wang, Yongcai Wang et al.

Collaborative 3D object detection holds significant importance in the field of autonomous driving, as it greatly enhances the perception capabilities of each individual agent by facilitating information exchange among multiple agents. However, in practice, due to pose estimation errors and time delays, the fusion of information across agents often results in feature representations with spatial and temporal noise, leading to detection errors. Diffusion models naturally have the ability to denoise noisy samples to the ideal data, which motivates us to explore the use of diffusion models to address the noise problem between multi-agent systems. In this work, we propose CoDiff, a novel robust collaborative perception framework that leverages the potential of diffusion models to generate more comprehensive and clearer feature representations. To the best of our knowledge, this is the first work to apply diffusion models to multi-agent collaborative perception. Specifically, we project high-dimensional feature map into the latent space of a powerful pre-trained autoencoder. Within this space, individual agent information serves as a condition to guide the diffusion model's sampling. This process denoises coarse feature maps and progressively refines the fused features. Experimental study on both simulated and real-world datasets demonstrates that the proposed framework CoDiff consistently outperforms existing relevant methods in terms of the collaborative object detection performance, and exhibits highly desired robustness when the pose and delay information of agents is with high-level noise. The code is released at https://github.com/HuangZhe885/CoDiff

Qingsong Xie, Zhao Zhang, Zhe Huang et al.

Image tokenization has significantly advanced visual generation and multimodal modeling, particularly when paired with autoregressive models. However, current methods face challenges in balancing efficiency and fidelity: high-resolution image reconstruction either requires an excessive number of tokens or compromises critical details through token reduction. To resolve this, we propose Latent Consistency Tokenizer (Layton) that bridges discrete visual tokens with the compact latent space of pre-trained Latent Diffusion Models (LDMs), enabling efficient representation of 1024x1024 images using only 256 tokens-a 16 times compression over VQGAN. Layton integrates a transformer encoder, a quantized codebook, and a latent consistency decoder. Direct application of LDM as the decoder results in color and brightness discrepancies. Thus, we convert it to latent consistency decoder, reducing multi-step sampling to 1-2 steps for direct pixel-level supervision. Experiments demonstrate Layton's superiority in high-fidelity reconstruction, with 10.8 reconstruction Frechet Inception Distance on MSCOCO-2017 5K benchmark for 1024x1024 image reconstruction. We also extend Layton to a text-to-image generation model, LaytonGen, working in autoregression. It achieves 0.73 score on GenEval benchmark, surpassing current state-of-the-art methods. Project homepage: https://github.com/OPPO-Mente-Lab/Layton

Zhe Huang, Ruohua Li, Kazuki Shin et al.

Multi-pedestrian trajectory prediction is an indispensable element of autonomous systems that safely interact with crowds in unstructured environments. Many recent efforts in trajectory prediction algorithms have focused on understanding social norms behind pedestrian motions. Yet we observe these works usually hold two assumptions, which prevent them from being smoothly applied to robot applications: (1) positions of all pedestrians are consistently tracked, and (2) the target agent pays attention to all pedestrians in the scene. The first assumption leads to biased interaction modeling with incomplete pedestrian data. The second assumption introduces aggregation of redundant surrounding information, and the target agent may be affected by unimportant neighbors or present overly conservative motion. Thus, we propose Gumbel Social Transformer, in which an Edge Gumbel Selector samples a sparse interaction graph of partially detected pedestrians at each time step. A Node Transformer Encoder and a Masked LSTM encode pedestrian features with sampled sparse graphs to predict trajectories. We demonstrate that our model overcomes potential problems caused by the aforementioned assumptions, and our approach outperforms related works in trajectory prediction benchmarks. Code is available at \url{https://github.com/tedhuang96/gst}.

Zhe Huang, Aamir Hasan, Kazuki Shin et al.

Trajectory prediction is one of the key capabilities for robots to safely navigate and interact with pedestrians. Critical insights from human intention and behavioral patterns need to be integrated to effectively forecast long-term pedestrian behavior. Thus, we propose a framework incorporating a Mutable Intention Filter and a Warp LSTM (MIF-WLSTM) to simultaneously estimate human intention and perform trajectory prediction. The Mutable Intention Filter is inspired by particle filtering and genetic algorithms, where particles represent intention hypotheses that can be mutated throughout the pedestrian motion. Instead of predicting sequential displacement over time, our Warp LSTM learns to generate offsets on a full trajectory predicted by a nominal intention-aware linear model, which considers the intention hypotheses during filtering process. Through experiments on a publicly available dataset, we show that our method outperforms baseline approaches and demonstrate the robust performance of our method under abnormal intention-changing scenarios. Code is available at https://github.com/tedhuang96/mifwlstm.

Guanchen Wu, Zhe Huang, Yuzhang Xie et al.

Reliable Alzheimer's disease (AD) diagnosis increasingly relies on multimodal assessments combining structural Magnetic Resonance Imaging (MRI) and Electronic Health Records (EHR). However, deploying these models is bottlenecked by modality missingness, as MRI scans are expensive and frequently unavailable in many patient cohorts. Furthermore, synthesizing de novo 3D anatomical scans from sparse, high-dimensional tabular records is technically challenging and poses severe clinical risks. To address this, we introduce MIRAGE, a novel framework that reframes the missing-MRI problem as an anatomy-guided cross-modal latent distillation task. First, MIRAGE leverages a Biomedical Knowledge Graph (KG) and Graph Attention Networks to map heterogeneous EHR variables into a unified embedding space that can be propagated from cohorts with real MRIs to cohorts without them. To bridge the semantic gap and enforce physical spatial awareness, we employ a frozen pre-trained 3D U-Net decoder strictly as an auxiliary regularization engine. Supported by a novel cohort-aggregated skip feature compensation strategy, this decoder acts as a rigorous structural penalty, forcing 1D latent representations to encode biologically plausible, macro-level pathological semantics. By exclusively utilizing this distilled "diagnostic-surrogate" representation during inference, MIRAGE completely bypasses computationally expensive 3D voxel reconstruction. Experiments demonstrate that our framework successfully bridges the missing-modality gap, improving the AD classification rate by 13% compared to unimodal baselines in cohorts without real MRIs.

Yiqian Gan, Hao Xiao, Yizhe Zhao et al.

Motion prediction has been an essential component of autonomous driving systems since it handles highly uncertain and complex scenarios involving moving agents of different types. In this paper, we propose a Multi-Granular TRansformer (MGTR) framework, an encoder-decoder network that exploits context features in different granularities for different kinds of traffic agents. To further enhance MGTR's capabilities, we leverage LiDAR point cloud data by incorporating LiDAR semantic features from an off-the-shelf LiDAR feature extractor. We evaluate MGTR on Waymo Open Dataset motion prediction benchmark and show that the proposed method achieved state-of-the-art performance, ranking 1st on its leaderboard (https://waymo.com/open/challenges/2023/motion-prediction/).

Xiaowei Yu, Zhe Huang, Zao Zhang

Unsupervised domain adaptation (UDA) aims to leverage the knowledge learned from labeled source domains to improve performance on the unlabeled target domains. While Convolutional Neural Networks (CNNs) have been dominant in previous UDA methods, recent research has shown promise in applying Vision Transformers (ViTs) to this task. In this study, we propose a novel Feature Fusion Transferability Aware Transformer (FFTAT) to enhance ViT performance in UDA tasks. Our method introduces two key innovations: First, we introduce a patch discriminator to evaluate the transferability of patches, generating a transferability matrix. We integrate this matrix into self-attention, directing the model to focus on transferable patches. Second, we propose a feature fusion technique to fuse embeddings in the latent space, enabling each embedding to incorporate information from all others, thereby improving generalization. These two components work in synergy to enhance feature representation learning. Extensive experiments on widely used benchmarks demonstrate that our method significantly improves UDA performance, achieving state-of-the-art (SOTA) results.

Zhe Huang, Xiaowei Yu, Benjamin S. Wessler et al.

Automated interpretation of ultrasound imaging of the heart (echocardiograms) could improve the detection and treatment of aortic stenosis (AS), a deadly heart disease. However, existing deep learning pipelines for assessing AS from echocardiograms have two key limitations. First, most methods rely on limited 2D cineloops, thereby ignoring widely available Doppler imaging that contains important complementary information about pressure gradients and blood flow abnormalities associated with AS. Second, obtaining labeled data is difficult. There are often far more unlabeled echocardiogram recordings available, but these remain underutilized by existing methods. To overcome these limitations, we introduce Semi-supervised Multimodal Multiple-Instance Learning (SMMIL), a new deep learning framework for automatic interpretation for structural heart diseases like AS. When deployed, SMMIL can combine information from two input modalities, spectral Dopplers and 2D cineloops, to produce a study-level AS diagnosis. During training, SMMIL can combine a smaller labeled set and an abundant unlabeled set of both modalities to improve its classifier. Experiments demonstrate that SMMIL outperforms recent alternatives at 3-level AS severity classification as well as several clinically relevant AS detection tasks.

Shuo Wang, Wanting Li, Yongcai Wang et al.

Deep visual odometry has demonstrated great advancements by learning-to-optimize technology. This approach heavily relies on the visual matching across frames. However, ambiguous matching in challenging scenarios leads to significant errors in geometric modeling and bundle adjustment optimization, which undermines the accuracy and robustness of pose estimation. To address this challenge, this paper proposes MambaVO, which conducts robust initialization, Mamba-based sequential matching refinement, and smoothed training to enhance the matching quality and improve the pose estimation. Specifically, the new frame is matched with the closest keyframe in the maintained Point-Frame Graph (PFG) via the semi-dense based Geometric Initialization Module (GIM). Then the initialized PFG is processed by a proposed Geometric Mamba Module (GMM), which exploits the matching features to refine the overall inter-frame matching. The refined PFG is finally processed by differentiable BA to optimize the poses and the map. To deal with the gradient variance, a Trending-Aware Penalty (TAP) is proposed to smooth training and enhance convergence and stability. A loop closure module is finally applied to enable MambaVO++. On public benchmarks, MambaVO and MambaVO++ demonstrate SOTA performance, while ensuring real-time running.

Zhe Huang, Yizhe Zhao, Hao Xiao et al.

Multi-view camera-only 3D object detection largely follows two primary paradigms: exploiting bird's-eye-view (BEV) representations or focusing on perspective-view (PV) features, each with distinct advantages. Although several recent approaches explore combining BEV and PV, many rely on partial fusion or maintain separate detection heads. In this paper, we propose DuoSpaceNet, a novel framework that fully unifies BEV and PV feature spaces within a single detection pipeline for comprehensive 3D perception. Our design includes a decoder to integrate BEV and PV features into unified detection queries, as well as a feature enhancement strategy that enriches different feature representations. In addition, DuoSpaceNet can be extended to handle multi-frame inputs, enabling more robust temporal analysis. Extensive experiments on nuScenes dataset show that DuoSpaceNet surpasses both BEV-based baselines (e.g., BEVFormer) and PV-based baselines (e.g., Sparse4D) in 3D object detection and BEV map segmentation, verifying the effectiveness of our proposed design.

Zhe Huang, Simone Rossi, Rui Yuan et al.

Transformers have become a standard architecture in machine learning, demonstrating strong in-context learning (ICL) abilities that allow them to learn from the prompt at inference time. However, uncertainty quantification for ICL remains an open challenge, particularly in noisy regression tasks. This paper investigates whether ICL can be leveraged for distribution-free uncertainty estimation, proposing a method based on conformal prediction to construct prediction intervals with guaranteed coverage. While traditional conformal methods are computationally expensive due to repeated model fitting, we exploit ICL to efficiently generate confidence intervals in a single forward pass. Our empirical analysis compares this approach against ridge regression-based conformal methods, showing that conformal prediction with in-context learning (CP with ICL) achieves robust and scalable uncertainty estimates. Additionally, we evaluate its performance under distribution shifts and establish scaling laws to guide model training. These findings bridge ICL and conformal prediction, providing a theoretically grounded and new framework for uncertainty quantification in transformer-based models.

Fangzhou Lin, Guoshun He, Zhenyu Guo et al.

Electric load is simultaneously affected across multiple time scales by exogenous factors such as weather and calendar rhythms, sudden events, and policies. Therefore, this paper proposes GRAFT (GRid-Aware Forecasting with Text), which modifies and improves STanHOP to better support grid-aware forecasting and multi-source textual interventions. Specifically, GRAFT strictly aligns daily-aggregated news, social media, and policy texts with half-hour load, and realizes text-guided fusion to specific time positions via cross-attention during both training and rolling forecasting. In addition, GRAFT provides a plug-and-play external-memory interface to accommodate different information sources in real-world deployment. We construct and release a unified aligned benchmark covering 2019--2021 for five Australian states (half-hour load, daily-aligned weather/calendar variables, and three categories of external texts), and conduct systematic, reproducible evaluations at three scales -- hourly, daily, and monthly -- under a unified protocol for comparison across regions, external sources, and time scales. Experimental results show that GRAFT significantly outperforms strong baselines and reaches or surpasses the state of the art across multiple regions and forecasting horizons. Moreover, the model is robust in event-driven scenarios and enables temporal localization and source-level interpretation of text-to-load effects through attention read-out. We release the benchmark, preprocessing scripts, and forecasting results to facilitate standardized empirical evaluation and reproducibility in power grid load forecasting.

Guorui Song, Guocun Wang, Zhe Huang et al.

Generating accurate descriptions of human actions in videos remains a challenging task for video captioning models. Existing approaches often struggle to capture fine-grained motion details, resulting in vague or semantically inconsistent captions. In this work, we introduce the Motion-Augmented Caption Model (M-ACM), a novel generative framework that enhances caption quality by incorporating motion-aware decoding. At its core, M-ACM leverages motion representations derived from human mesh recovery to explicitly highlight human body dynamics, thereby reducing hallucinations and improving both semantic fidelity and spatial alignment in the generated captions. To support research in this area, we present the Human Motion Insight (HMI) Dataset, comprising 115K video-description pairs focused on human movement, along with HMI-Bench, a dedicated benchmark for evaluating motion-focused video captioning. Experimental results demonstrate that M-ACM significantly outperforms previous methods in accurately describing complex human motions and subtle temporal variations, setting a new standard for motion-centric video captioning.

Zhe Huang, John Pohovey, Ananya Yammanuru et al.

Large Language Models (LLM) and Vision Language Models (VLM) enable robots to ground natural language prompts into control actions to achieve tasks in an open world. However, when applied to a long-horizon collaborative task, this formulation results in excessive prompting for initiating or clarifying robot actions at every step of the task. We propose Language-driven Intention Tracking (LIT), leveraging LLMs and VLMs to model the human user's long-term behavior and to predict the next human intention to guide the robot for proactive collaboration. We demonstrate smooth coordination between a LIT-based collaborative robot and the human user in collaborative cooking tasks.

Zhe Huang, Xiaowei Yu, Dajiang Zhu et al.

Semi-supervised learning (SSL) seeks to enhance task performance by training on both labeled and unlabeled data. Mainstream SSL image classification methods mostly optimize a loss that additively combines a supervised classification objective with a regularization term derived solely from unlabeled data. This formulation neglects the potential for interaction between labeled and unlabeled images. In this paper, we introduce InterLUDE, a new approach to enhance SSL made of two parts that each benefit from labeled-unlabeled interaction. The first part, embedding fusion, interpolates between labeled and unlabeled embeddings to improve representation learning. The second part is a new loss, grounded in the principle of consistency regularization, that aims to minimize discrepancies in the model's predictions between labeled versus unlabeled inputs. Experiments on standard closed-set SSL benchmarks and a medical SSL task with an uncurated unlabeled set show clear benefits to our approach. On the STL-10 dataset with only 40 labels, InterLUDE achieves 3.2% error rate, while the best previous method reports 14.9%.

Zhe Huang, Yudian Li

Most generic object detectors are mainly built for standard object detection tasks such as COCO and PASCAL VOC. They might not work well and/or efficiently on tasks of other domains consisting of images that are visually different from standard datasets. To this end, many advances have been focused on adapting a general-purposed object detector with limited domain-specific designs. However, designing a successful task-specific detector requires extraneous manual experiments and parameter tuning through trial and error. In this paper, we first propose and examine a fully-automatic pipeline to design a fully-specialized detector (FSD) which mainly incorporates a neural-architectural-searched model by exploring ideal network structures over the backbone and task-specific head. On the DeepLesion dataset, extensive results show that FSD can achieve 3.1 mAP gain while using approximately 40% fewer parameters on binary lesion detection task and improved the mAP by around 10% on multi-type lesion detection task via our region-aware graph modeling compared with existing general-purposed medical lesion detection networks.

Zhe Huang, Benjamin S. Wessler, Michael C. Hughes

Aortic stenosis (AS) is a degenerative valve condition that causes substantial morbidity and mortality. This condition is under-diagnosed and under-treated. In clinical practice, AS is diagnosed with expert review of transthoracic echocardiography, which produces dozens of ultrasound images of the heart. Only some of these views show the aortic valve. To automate screening for AS, deep networks must learn to mimic a human expert's ability to identify views of the aortic valve then aggregate across these relevant images to produce a study-level diagnosis. We find previous approaches to AS detection yield insufficient accuracy due to relying on inflexible averages across images. We further find that off-the-shelf attention-based multiple instance learning (MIL) performs poorly. We contribute a new end-to-end MIL approach with two key methodological innovations. First, a supervised attention technique guides the learned attention mechanism to favor relevant views. Second, a novel self-supervised pretraining strategy applies contrastive learning on the representation of the whole study instead of individual images as commonly done in prior literature. Experiments on an open-access dataset and an external validation set show that our approach yields higher accuracy while reducing model size.

Zhe Huang, Gary Long, Benjamin Wessler et al.

Semi-supervised image classification has shown substantial progress in learning from limited labeled data, but recent advances remain largely untested for clinical applications. Motivated by the urgent need to improve timely diagnosis of life-threatening heart conditions, especially aortic stenosis, we develop a benchmark dataset to assess semi-supervised approaches to two tasks relevant to cardiac ultrasound (echocardiogram) interpretation: view classification and disease severity classification. We find that a state-of-the-art method called MixMatch achieves promising gains in heldout accuracy on both tasks, learning from a large volume of truly unlabeled images as well as a labeled set collected at great expense to achieve better performance than is possible with the labeled set alone. We further pursue patient-level diagnosis prediction, which requires aggregating across hundreds of images of diverse view types, most of which are irrelevant, to make a coherent prediction. The best patient-level performance is achieved by new methods that prioritize diagnosis predictions from images that are predicted to be clinically-relevant views and transfer knowledge from the view task to the diagnosis task. We hope our released Tufts Medical Echocardiogram Dataset and evaluation framework inspire further improvements in multi-task semi-supervised learning for clinical applications.

Nan Wu, Zhe Huang, Yiqiu Shen et al.

Breast cancer is the most common cancer in women, and hundreds of thousands of unnecessary biopsies are done around the world at a tremendous cost. It is crucial to reduce the rate of biopsies that turn out to be benign tissue. In this study, we build deep neural networks (DNNs) to classify biopsied lesions as being either malignant or benign, with the goal of using these networks as second readers serving radiologists to further reduce the number of false positive findings. We enhance the performance of DNNs that are trained to learn from small image patches by integrating global context provided in the form of saliency maps learned from the entire image into their reasoning, similar to how radiologists consider global context when evaluating areas of interest. Our experiments are conducted on a dataset of 229,426 screening mammography exams from 141,473 patients. We achieve an AUC of 0.8 on a test set consisting of 464 benign and 136 malignant lesions.

Peter Du, Zhe Huang, Tianqi Liu et al.

As autonomous systems begin to operate amongst humans, methods for safe interaction must be investigated. We consider an example of a small autonomous vehicle in a pedestrian zone that must safely maneuver around people in a free-form fashion. We investigate two key questions: How can we effectively integrate pedestrian intent estimation into our autonomous stack. Can we develop an online monitoring framework to give formal guarantees on the safety of such human-robot interactions. We present a pedestrian intent estimation framework that can accurately predict future pedestrian trajectories given multiple possible goal locations. We integrate this into a reachability-based online monitoring scheme that formally assesses the safety of these interactions with nearly real-time performance (approximately 0.3 seconds). These techniques are integrated on a test vehicle with a complete in-house autonomous stack, demonstrating effective and safe interaction in real-world experiments.

Zhe Huang, Weijiang Yu, Wayne Zhang et al.

Most advances in single image de-raining meet a key challenge, which is removing rain streaks with different scales and shapes while preserving image details. Existing single image de-raining approaches treat rain-streak removal as a process of pixel-wise regression directly. However, they are lacking in mining the balance between over-de-raining (e.g. removing texture details in rain-free regions) and under-de-raining (e.g. leaving rain streaks). In this paper, we firstly propose a coarse-to-fine network called Gradual Network (GraNet) consisting of coarse stage and fine stage for delving into single image de-raining with different granularities. Specifically, to reveal coarse-grained rain-streak characteristics (e.g. long and thick rain streaks/raindrops), we propose a coarse stage by utilizing local-global spatial dependencies via a local-global subnetwork composed of region-aware blocks. Taking the residual result (the coarse de-rained result) between the rainy image sample (i.e. the input data) and the output of coarse stage (i.e. the learnt rain mask) as input, the fine stage continues to de-rain by removing the fine-grained rain streaks (e.g. light rain streaks and water mist) to get a rain-free and well-reconstructed output image via a unified contextual merging sub-network with dense blocks and a merging block. Solid and comprehensive experiments on synthetic and real data demonstrate that our GraNet can significantly outperform the state-of-the-art methods by removing rain streaks with various densities, scales and shapes while keeping the image details of rain-free regions well-preserved.

Nan Wu, Jason Phang, Jungkyu Park et al.

We present a deep convolutional neural network for breast cancer screening exam classification, trained and evaluated on over 200,000 exams (over 1,000,000 images). Our network achieves an AUC of 0.895 in predicting whether there is a cancer in the breast, when tested on the screening population. We attribute the high accuracy of our model to a two-stage training procedure, which allows us to use a very high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and find our model to be as accurate as experienced radiologists when presented with the same data. Finally, we show that a hybrid model, averaging probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To better understand our results, we conduct a thorough analysis of our network's performance on different subpopulations of the screening population, model design, training procedure, errors, and properties of its internal representations.