Abstract:Unsupervised Domain Adaptation (UDA) aims to align source and target domain distributions to close the domain gap, but still struggles with obtaining the target data. Fortunately, Domain Generalization (DG) excels without the need for any target data. Recent works expose that depth maps contribute to improved generalized performance in the UDA tasks, but they ignore the noise and holes in depth maps due to device and environmental factors, failing to sufficiently and effectively learn domain-invariant representation. Although high-sensitivity region suppression has shown promising results in learning domain-invariant features, existing methods cannot be directly applicable to depth maps due to their unique characteristics. Hence, we propose a novel framework, namely Depth-Sensitive Soft Suppression with RGB-D inter-modal stylization flow (DSSS), focusing on learning domain-invariant features from depth maps for the DG semantic segmentation. Specifically, we propose the RGB-D inter-modal stylization flow to generate stylized depth maps for sensitivity detection, cleverly utilizing RGB information as the stylization source. Then, a class-wise soft spatial sensitivity suppression is designed to identify and emphasize non-sensitive depth features that contain more domain-invariant information. Furthermore, an RGB-D soft alignment loss is proposed to ensure that the stylized depth maps only align part of the RGB features while still retaining the unique depth information. To our best knowledge, our DSSS framework is the first work to integrate RGB and Depth information in the multi-class DG semantic segmentation task. Extensive experiments over multiple backbone networks show that our framework achieves remarkable performance improvement.
Abstract:Autonomous driving has made significant strides through data-driven techniques, achieving robust performance in standardized tasks. However, existing methods frequently overlook user-specific preferences, offering limited scope for interaction and adaptation with users. To address these challenges, we propose a "fast-slow" decision-making framework that integrates a Large Language Model (LLM) for high-level instruction parsing with a Reinforcement Learning (RL) agent for low-level real-time decision. In this dual system, the LLM operates as the "slow" module, translating user directives into structured guidance, while the RL agent functions as the "fast" module, making time-critical maneuvers under stringent latency constraints. By decoupling high-level decision making from rapid control, our framework enables personalized user-centric operation while maintaining robust safety margins. Experimental evaluations across various driving scenarios demonstrate the effectiveness of our method. Compared to baseline algorithms, the proposed architecture not only reduces collision rates but also aligns driving behaviors more closely with user preferences, thereby achieving a human-centric mode. By integrating user guidance at the decision level and refining it with real-time control, our framework bridges the gap between individual passenger needs and the rigor required for safe, reliable driving in complex traffic environments.
Abstract:Gaussian Splatting (GS) and Neural Radiance Fields (NeRF) are two groundbreaking technologies that have revolutionized the field of Novel View Synthesis (NVS), enabling immersive photorealistic rendering and user experiences by synthesizing multiple viewpoints from a set of images of sparse views. The potential applications of NVS, such as high-quality virtual and augmented reality, detailed 3D modeling, and realistic medical organ imaging, underscore the importance of quality assessment of NVS methods from the perspective of human perception. Although some previous studies have explored subjective quality assessments for NVS technology, they still face several challenges, especially in NVS methods selection, scenario coverage, and evaluation methodology. To address these challenges, we conducted two subjective experiments for the quality assessment of NVS technologies containing both GS-based and NeRF-based methods, focusing on dynamic and real-world scenes. This study covers 360{\deg}, front-facing, and single-viewpoint videos while providing a richer and greater number of real scenes. Meanwhile, it's the first time to explore the impact of NVS methods in dynamic scenes with moving objects. The two types of subjective experiments help to fully comprehend the influences of different viewing paths from a human perception perspective and pave the way for future development of full-reference and no-reference quality metrics. In addition, we established a comprehensive benchmark of various state-of-the-art objective metrics on the proposed database, highlighting that existing methods still struggle to accurately capture subjective quality. The results give us some insights into the limitations of existing NVS methods and may promote the development of new NVS methods.
Abstract:Faces and humans are crucial elements in social interaction and are widely included in everyday photos and videos. Therefore, a deep understanding of faces and humans will enable multi-modal assistants to achieve improved response quality and broadened application scope. Currently, the multi-modal assistant community lacks a comprehensive and scientific evaluation of face and human understanding abilities. In this paper, we first propose a hierarchical ability taxonomy that includes three levels of abilities. Then, based on this taxonomy, we collect images and annotations from publicly available datasets in the face and human community and build a semi-automatic data pipeline to produce problems for the new benchmark. Finally, the obtained Face-Human-Bench comprises a development set with 900 problems and a test set with 1800 problems, supporting both English and Chinese. We conduct evaluations over 25 mainstream multi-modal large language models (MLLMs) with our Face-Human-Bench, focusing on the correlation between abilities, the impact of the relative position of targets on performance, and the impact of Chain of Thought (CoT) prompting on performance. Moreover, inspired by multi-modal agents, we also explore which abilities of MLLMs need to be supplemented by specialist models.
Abstract:Current methods for generating human motion videos rely on extracting pose sequences from reference videos, which restricts flexibility and control. Additionally, due to the limitations of pose detection techniques, the extracted pose sequences can sometimes be inaccurate, leading to low-quality video outputs. We introduce a novel task aimed at generating human motion videos solely from reference images and natural language. This approach offers greater flexibility and ease of use, as text is more accessible than the desired guidance videos. However, training an end-to-end model for this task requires millions of high-quality text and human motion video pairs, which are challenging to obtain. To address this, we propose a new framework called Fleximo, which leverages large-scale pre-trained text-to-3D motion models. This approach is not straightforward, as the text-generated skeletons may not consistently match the scale of the reference image and may lack detailed information. To overcome these challenges, we introduce an anchor point based rescale method and design a skeleton adapter to fill in missing details and bridge the gap between text-to-motion and motion-to-video generation. We also propose a video refinement process to further enhance video quality. A large language model (LLM) is employed to decompose natural language into discrete motion sequences, enabling the generation of motion videos of any desired length. To assess the performance of Fleximo, we introduce a new benchmark called MotionBench, which includes 400 videos across 20 identities and 20 motions. We also propose a new metric, MotionScore, to evaluate the accuracy of motion following. Both qualitative and quantitative results demonstrate that our method outperforms existing text-conditioned image-to-video generation methods. All code and model weights will be made publicly available.
Abstract:Federated learning, a novel paradigm designed to protect data privacy, is vulnerable to backdoor attacks due to its distributed nature. Current research often designs attacks based on a single attacker with a single backdoor, overlooking more realistic and complex threats in federated learning. We propose a more practical threat model for federated learning: the distributed multi-target backdoor. In this model, multiple attackers control different clients, embedding various triggers and targeting different classes, collaboratively implanting backdoors into the global model via central aggregation. Empirical validation shows that existing methods struggle to maintain the effectiveness of multiple backdoors in the global model. Our key insight is that similar backdoor triggers cause parameter conflicts and injecting new backdoors disrupts gradient directions, significantly weakening some backdoors performance. To solve this, we propose a Distributed Multi-Target Backdoor Attack (DMBA), ensuring efficiency and persistence of backdoors from different malicious clients. To avoid parameter conflicts, we design a multi-channel dispersed frequency trigger strategy to maximize trigger differences. To mitigate gradient interference, we introduce backdoor replay in local training to neutralize conflicting gradients. Extensive validation shows that 30 rounds after the attack, Attack Success Rates of three different backdoors from various clients remain above 93%. The code will be made publicly available after the review period.
Abstract:Adapting large pre-trained foundation models, e.g., SAM, for medical image segmentation remains a significant challenge. A crucial step involves the formulation of a series of specialized prompts that incorporate specific clinical instructions. Past works have been heavily reliant on a singular type of prompt for each instance, necessitating manual input of an ideally correct prompt, which is less efficient. To tackle this issue, we propose to utilize prompts of different granularity, which are sourced from original images to provide a broader scope of clinical insights. However, combining prompts of varying types can pose a challenge due to potential conflicts. In response, we have designed a coarse-to-fine mechanism, referred to as curriculum prompting, that progressively integrates prompts of different types. Through extensive experiments on three public medical datasets across various modalities, we demonstrate the effectiveness of our proposed approach, which not only automates the prompt generation process but also yields superior performance compared to other SAM-based medical image segmentation methods. Code is available at: https://github.com/AnnaZzz-zxq/Curriculum-Prompting.
Abstract:SOTA facial expression recognition (FER) methods fail on test sets that have domain gaps with the train set. Recent domain adaptation FER methods need to acquire labeled or unlabeled samples of target domains to fine-tune the FER model, which might be infeasible in real-world deployment. In this paper, we aim to improve the zero-shot generalization ability of FER methods on different unseen test sets using only one train set. Inspired by how humans first detect faces and then select expression features, we propose a novel FER pipeline to extract expression-related features from any given face images. Our method is based on the generalizable face features extracted by large models like CLIP. However, it is non-trivial to adapt the general features of CLIP for specific tasks like FER. To preserve the generalization ability of CLIP and the high precision of the FER model, we design a novel approach that learns sigmoid masks based on the fixed CLIP face features to extract expression features. To further improve the generalization ability on unseen test sets, we separate the channels of the learned masked features according to the expression classes to directly generate logits and avoid using the FC layer to reduce overfitting. We also introduce a channel-diverse loss to make the learned masks separated. Extensive experiments on five different FER datasets verify that our method outperforms SOTA FER methods by large margins. Code is available in https://github.com/zyh-uaiaaaa/Generalizable-FER.
Abstract:Story visualization aims to create visually compelling images or videos corresponding to textual narratives. Despite recent advances in diffusion models yielding promising results, existing methods still struggle to create a coherent sequence of subject-consistent frames based solely on a story. To this end, we propose DreamStory, an automatic open-domain story visualization framework by leveraging the LLMs and a novel multi-subject consistent diffusion model. DreamStory consists of (1) an LLM acting as a story director and (2) an innovative Multi-Subject consistent Diffusion model (MSD) for generating consistent multi-subject across the images. First, DreamStory employs the LLM to generate descriptive prompts for subjects and scenes aligned with the story, annotating each scene's subjects for subsequent subject-consistent generation. Second, DreamStory utilizes these detailed subject descriptions to create portraits of the subjects, with these portraits and their corresponding textual information serving as multimodal anchors (guidance). Finally, the MSD uses these multimodal anchors to generate story scenes with consistent multi-subject. Specifically, the MSD includes Masked Mutual Self-Attention (MMSA) and Masked Mutual Cross-Attention (MMCA) modules. MMSA and MMCA modules ensure appearance and semantic consistency with reference images and text, respectively. Both modules employ masking mechanisms to prevent subject blending. To validate our approach and promote progress in story visualization, we established a benchmark, DS-500, which can assess the overall performance of the story visualization framework, subject-identification accuracy, and the consistency of the generation model. Extensive experiments validate the effectiveness of DreamStory in both subjective and objective evaluations. Please visit our project homepage at https://dream-xyz.github.io/dreamstory.
Abstract:Early and accurate detection of anomalous events on the freeway, such as accidents, can improve emergency response and clearance. However, existing delays and errors in event identification and reporting make it a difficult problem to solve. Current large-scale freeway traffic datasets are not designed for anomaly detection and ignore these challenges. In this paper, we introduce the first large-scale lane-level freeway traffic dataset for anomaly detection. Our dataset consists of a month of weekday radar detection sensor data collected in 4 lanes along an 18-mile stretch of Interstate 24 heading toward Nashville, TN, comprising over 3.7 million sensor measurements. We also collect official crash reports from the Nashville Traffic Management Center and manually label all other potential anomalies in the dataset. To show the potential for our dataset to be used in future machine learning and traffic research, we benchmark numerous deep learning anomaly detection models on our dataset. We find that unsupervised graph neural network autoencoders are a promising solution for this problem and that ignoring spatial relationships leads to decreased performance. We demonstrate that our methods can reduce reporting delays by over 10 minutes on average while detecting 75% of crashes. Our dataset and all preprocessing code needed to get started are publicly released at https://vu.edu/ft-aed/ to facilitate future research.