AAAI 2025 Papers

Traffic classification is crucial for network management and security. Recently, deep learning-based methods have demonstrated good performance in traffic classification. However, they primarily capture features from raw packet bytes, overlooking the significance of inter-packet correlations within flows from a global perspective. Additionally, effectively handling both packet-length and temporal information, while extracting the structural relationships from a graph into the model, remains a challenge for enhancing the performance of traffic prediction. In this paper, we propose DigTraffic, a novel dual-channel interactive graph transformer to address these limitations. DigTraffic employs a message-level graph-structured flow representation combined with message-aware structural aggregation. To learn intrinsic flow representations, DigTraffic constructs traffic interaction graphs, by incorporating three well-designed heterogeneous types of edges to capture client-server interactions. After that, we separately encode lengthy and temporal flow sequences using a dual-channel network and fuse these modalities within a Transformer architecture. Furthermore, DigTraffic introduces a message-aware Graph Transformer that leverages both node embeddings and edge spatial relations to capture complex graph structures and rich structural information. Experimental results demonstrate that our method significantly outperforms the state-of-the-art methods on four real-world traffic datasets.

With the rapid advancement of 3D scanning technology, point clouds have become a crucial data type in computer vision and machine learning. However, learning robust representations for point clouds remains a significant challenge due to their irregularity and sparsity. In this paper, we propose a novel Dual Manifold Regularization (DMR) framework that makes full use of the properties of positive and negative curvature in manifolds to improve the representation of point clouds. Specifically, we leverage DMR based on hyperbolic and hyperspherical manifolds to address the limitations of traditional single-manifold regularization techniques, including inadequate generalization ability and adaptability to data diversity, as well as the difficulty of capturing complex relationships between data. To begin, we utilize the tree-like structure of the hyperbolic manifold to model the part-whole hierarchical relationships within point clouds. This allows for a more comprehensive representation of the data, improving the model's capability to understand complex shapes. Additionally, we construct positive samples through topological consistency augmentation and employ contrastive learning techniques in the hyperspherical manifold to capture more discriminative features within the data. Our experimental results show that our method outperforms traditional supervised learning and single-manifold regularization techniques in point cloud analysis. Specifically, for shape classification, DMR achieves a new State-Of-The-Art (SOTA) performance with 94.8% Overall Accuracy (OA) on ModelNet40 and 90.7% OA on ScanObjectNN, surpassing the recent SOTA model without increasing the baseline parameters.

Synthetic aperture radar (SAR) object detection requires accurate identification and localization of targets at various scales within SAR images. However, background clutter and speckle noise can obscure key features and mislead the knowledge distillation process. To address these challenges, we introduce the Dual Information Purification Knowledge Distillation (DIPKD) method, which improves the performance of the student model through three key strategies: denoising, enrichment, and decoupling. First, our Selective Noise Suppression (SNS) technique reduces speckle noise in global features by minimizing misleading information from the teacher model. Second, the Knowledge Level Decoupling (KLD) module separates features into target and non-target knowledge, balancing feature mapping and reducing background noise to enhance the extraction of critical information for the student model. Finally, the Reverse Information Transfer (RIT) module refines intermediate features in the student model, compensating for the loss of detailed local information. Experimental results demonstrate that DIPKD significantly outperforms existing distillation techniques in SAR object detection, achieving 60.2% and 51.4% mAP scores on the SSDD and HRSID datasets, respectively. Additionally, the student model shows performance improvements of 1.3% and 2.9% over the teacher model, highlighting the effectiveness of the information purification approach.

Video Anomaly Detection (VAD) is essential for computer vision and multimedia research. Existing VAD methods utilize either reconstruction-based or prediction-based frameworks. The former excels at detecting irregular patterns or structures, whereas the latter is capable of spotting abnormal deviations or trends. We address pose-based video anomaly detection and introduce a novel framework called Dual Conditioned Motion Diffusion (DCMD), which enjoys the advantages of both approaches. The DCMD integrates conditioned motion and conditioned embedding to comprehensively utilize the pose characteristics and latent semantics of observed movements, respectively. In the reverse diffusion process, a motion transformer is proposed to capture potential correlations from multi-layered characteristics within the spectrum space of human motion. To enhance the discriminability between normal and abnormal instances, we design a novel United Association Discrepancy (UAD) regularization that primarily relies on a Gaussian kernel-based time association and a self-attention-based global association. Finally, a mask completion strategy is introduced during the inference stage of the reverse diffusion process to enhance the utilization of conditioned motion for the prediction branch of anomaly detection. Extensive experiments conducted on four datasets demonstrate that our method dramatically outperforms state-of-the-art methods and exhibits superior generalization performance.

Semi-supervised medical image segmentation (SSMIS) uses consistency learning to regularize model training, which alleviates the burden of pixel-wise manual annotations. However, it often suffers from error supervision from low-quality pseudo labels. Vision-Language Model (VLM) has great potential to enhance pseudo labels by introducing text prompt guided multimodal supervision information. It nevertheless faces the cross-modal problem: the obtained messages tend to correspond to multiple targets. To address aforementioned problems, we propose a Dual Semantic Similarity-Supervised VLM (DuSSS) for SSMIS. Specifically, 1) a Dual Contrastive Learning (DCL) is designed to improve cross-modal semantic consistency by capturing intrinsic representations within each modality and semantic correlations across modalities. 2) To encourage the learning of multiple semantic correspondences, a Semantic Similarity-Supervision strategy (SSS) is proposed and injected into each contrastive learning process in DCL, supervising semantic similarity via the distribution-based uncertainty levels. Furthermore, a novel VLM-based SSMIS network is designed to compensate for the quality deficiencies of pseudo-labels. It utilizes the pretrained VLM to generate text prompt guided supervision information, refining the pseudo label for better consistency regularization. Experimental results demonstrate that our DuSSS achieves outstanding performance with Dice of 82.52%, 74.61% and 78.03% on three public datasets (QaTa-COV19, BM-Seg and MoNuSeg).

The exceptional generative capability of text-to-image models has raised substantial safety concerns regarding the generation of Not-Safe-For-Work (NSFW) content and potential copyright infringement. To address these concerns, previous methods safeguard the models by eliminating inappropriate concepts. Nonetheless, these models alter the parameters of the backbone network and exert considerable influences on the structural (low-frequency) components of the image, which undermines the model's ability to retain irrelevant concepts. In this work, we propose our Dual encoder Modulation network (DuMo), which achieves precise erasure of inappropriate target concepts with minimum impairment to non-target concepts. In contrast to previous methods, DuMo employs the Eraser with PRior Knowledge (EPR) module which modifies the skip connection features of the U-NET and primarily achieves concept erasure on details (high-frequency) components of the image. To minimize the demage to non-target concepts during erasure, the parameters of the backbone U-NET are frozen and the prior knowledge from the original skip connection features is introduced to the erasure process. Meanwhile, the phenomenon is observed that distinct erasing preferences for the image structure and details are demonstrated by the EPR at different timesteps and layers. Therefore, we adopt a novel Time-Layer MOdulation process (TLMO) that adjusts the erasure scale of EPR module's outputs across different layers and timesteps, automatically balancing the erasure effects and model's generative ability. Our method achieves state-of-the-art performance on Explicit Content Erasure (detecting only 34 nude parts), Cartoon Concept Removal (with an average LPIPS_da of 0.428, 0.113 higher than SOTA at 0.315), and Artistic Style Erasure (with an average LPIPS_da of 0.387, 0.088 higher than SOTA at 0.299), clearly outperforming alternative methods.

Virtual screening (VS) is a critical step in computer-aided drug discovery, aiming to identify molecules that bind to a specific target protein. Traditional VS methods, such as docking, are often too time-consuming to efficiently screen large-scale molecular databases. Recent advances in deep learning have demonstrated that learning vector representations for both proteins and molecules using contrastive learning can outperform traditional docking methods. However, considering that the target databases often contain billions of molecules, real-valued vector representations adopted by existing methods can still incur large memory and time cost in VS. To address this problem, we propose DrugHash, a hashing-based contrastive learning method for VS. DrugHash formulates VS as a retrieval task that leverages binary hash codes for efficient retrieval. In particular, DrugHash designs a simple yet effective hashing strategy to enable end-to-end learning of binary hash codes for both proteins and molecules, which can dramatically reduce the memory and time cost with higher accuracy compared with existing methods. Experimental results show that DrugHash can outperform existing methods to achieve state-of-the-art accuracy, with at least a 32 times reduction in memory cost and a 4.6 times improvement in speed.

Drought is one of the most destructive and expensive natural disasters, severely impacting natural resources and risks by depleting water resources and diminishing agricultural yields. Under climate change, accurately predicting drought is critical for mitigating drought-induced risks. However, the intricate interplay among the physical and biological drivers that regulate droughts limits the predictability and understanding of drought, particularly at a subseasonal to seasonal (S2S) time scale. While deep learning has demonstrated the potential to address climate forecasting challenges, its application to drought prediction has received relatively less attention. In this work, we propose a new dataset, DroughtSet, which integrates relevant predictive features and three drought indices from multiple remote sensing and reanalysis datasets across the contiguous United States (CONUS). DroughtSet specifically provides the machine learning community with a new real-world dataset to benchmark drought prediction models and more generally, time-series forecasting methods. Furthermore, we propose a spatial-temporal model SPDrought to predict and interpret S2S droughts. Our model learns from the spatial and temporal information of physical and biological features to predict three types of droughts simultaneously. Multiple strategies are employed to quantify the importance of physical and biological features for drought prediction. Our results provide insights for researchers to better understand the predictability and sensitivity of drought to biological and physical conditions. We aim to contribute to the climate field by proposing a new tool to predict and understand the occurrence of droughts and provide the AI community with a new benchmark to study deep learning applications in climate science.

Rap, a prominent genre of vocal performance, remains underexplored in vocal generation. General vocal synthesis depends on precise note and duration inputs, requiring users to have related musical knowledge, which limits flexibility. In contrast, rap typically features simpler melodies, with a core focus on a strong rhythmic sense that harmonizes with accompanying beats. In this paper, we propose Freestyler, the first system that generates rapping vocals directly from lyrics and accompaniment inputs. Freestyler utilizes language model-based token generation, followed by a conditional flow matching model to produce spectrograms and a neural vocoder to restore audio. It allows a 3-second prompt to enable zero-shot timbre control. Due to the scarcity of publicly available rap datasets, we also present RapBank, a rap song dataset collected from the internet, alongside a meticulously designed processing pipeline. Experimental results show that Freestyler produces high-quality rapping voice generation with enhanced naturalness and strong alignment with accompanying beats, both stylistically and rhythmically.

We propose DrivingForward, a feed-forward Gaussian Splatting model that reconstructs driving scenes from flexible surround-view input. Driving scene images from vehicle-mounted cameras are typically sparse, with limited overlap, and the movement of the vehicle further complicates the acquisition of camera extrinsics. To tackle these challenges and achieve real-time reconstruction, we jointly train a pose network, a depth network, and a Gaussian network to predict the Gaussian primitives that represent the driving scenes. The pose network and depth network determine the position of the Gaussian primitives in a self-supervised manner, without using depth ground truth and camera extrinsics during training. The Gaussian network independently predicts primitive parameters from each input image, including covariance, opacity, and spherical harmonics coefficients. At the inference stage, our model can achieve feed-forward reconstruction from flexible multi-frame surround-view input. Experiments on the nuScenes dataset show that our model outperforms existing state-of-the-art feed-forward and scene-optimized reconstruction methods in terms of reconstruction.

World models envision potential future states based on various ego actions. They embed extensive knowledge about the driving environment, facilitating safe and scalable autonomous driving. Most existing methods primarily focus on either data generation or the pretraining paradigms of world models. Unlike the aforementioned prior works, we propose Drive-OccWorld, which adapts a vision-centric 4D forecasting world model to end-to-end planning for autonomous driving. Specifically, we first introduce a semantic and motion-conditional normalization in the memory module, which accumulates semantic and dynamic information from historical BEV embeddings. These BEV features are then conveyed to the world decoder for future occupancy and flow forecasting, considering both geometry and spatiotemporal modeling. Additionally, we propose injecting flexible action conditions, such as velocity, steering angle, trajectory, and commands, into the world model to enable controllable generation and facilitate a broader range of downstream applications. Furthermore, we explore integrating the generative capabilities of the 4D world model with end-to-end planning, enabling continuous forecasting of future states and the selection of optimal trajectories using an occupancy-based cost function. Extensive experiments on the nuScenes dataset demonstrate that our method can generate plausible and controllable 4D occupancy, opening new avenues for driving world generation and end-to-end planning.

We introduce a wearable driving status recognition device and our open-source dataset, along with a new real-time method robust to changes in lighting conditions for identifying driving status from eye observations of drivers. The core of our method is generating event frames from conventional intensity frames, and the other is a newly designed Attention Driving State Network (ADSN). Compared to event cameras, conventional cameras offer complete information and lower hardware costs, enabling captured frames to encode rich spatial information. However, these textures lack temporal information, posing challenges in effectively identifying driving status. DriveGazen addresses this issue from three perspectives. First, we utilize video frames to generate realistic synthetic dynamic vision sensor (DVS) events.Second, we adopt a spiking neural network to decode pertinent temporal information. Lastly, ADSN extracts crucial spatial cues from corresponding intensity frames and conveys spatial attention to convolutional spiking layers during both training and inference through a novel guide attention module to guide the feature learning and feature enhancement of the event frame. We specifically collected the Driving Status (DriveGaze) dataset to demonstrate the effectiveness of our approach. Additionally, we validate the superiority of the DriveGazen on the Single-eye Event-based Emotion (SEE) dataset. To the best of our knowledge, our method is the first to utilize guide attention spiking neural networks and eye-based event frames generated from conventional cameras for driving status recognition.Please refer to our project page and supplementary materials for more details.

Vision-centric autonomous driving systems require diverse data for robust training and evaluation, which can be augmented by manipulating object positions and appearances within existing scene captures. While recent advancements in diffusion models have shown promise in video editing, their application to object manipulation in driving scenarios remains challenging due to imprecise positional control and difficulties in preserving high-fidelity object appearances. To address these challenges in position and appearance control, we introduce DriveEditor, the first diffusion-based framework for object editing in driving videos. DriveEditor offers a unified framework for comprehensive object editing operations, including repositioning, replacement, deletion, and insertion. These diverse manipulations are all achieved through a shared set of varying inputs, processed by identical position control and appearance maintenance modules. The position control module projects the given 3D bounding box while preserving depth information and hierarchically injects it into the diffusion process, enabling precise control over object position and orientation. The appearance maintenance module preserves consistent attributes with a single reference image by employing a three-tiered approach: low-level detail preservation, high-level semantic maintenance, and the integration of 3D priors from a novel view synthesis model. Extensive qualitative and quantitative evaluations on the nuScenes dataset demonstrate DriveEditor's exceptional fidelity and controllability in generating diverse driving scene edits, as well as its remarkable ability to facilitate downstream tasks.

World models have demonstrated superiority in autonomous driving, particularly in the generation of multi-view driving videos. However, significant challenges still exist in generating customized driving videos. In this paper, we propose DriveDreamer-2, which incorporates a Large Language Model (LLM) to facilitate the creation of user-defined driving videos. Specifically, a trajectory generation function library is developed to produce trajectories that conform to user descriptions. Subsequently, an HDMap generator is designed to learn the mapping from trajectories to road structures. Ultimately, we propose the Unified Multi-View Model (UniMVM) to enhance temporal and spatial coherence in the generated multi-view driving videos. To the best of our knowledge, DriveDreamer-2 is the first world model to generate customized driving videos, and it can generate uncommon driving videos (e.g., vehicles abruptly cut in) in a user-friendly manner. Besides, experimental results demonstrate that the generated videos enhance the training of driving perception methods (e.g., 3D detection and tracking). Furthermore, video generation quality of DriveDreamer-2 surpasses other state-of-the-art methods, showcasing FID and FVD scores of 11.2 and 55.7, representing relative improvements of ~30% and ~50%.

Existing ultra-high-definition (UHD) image restoration methods often struggle with consistency due to downsampling. We aim to address these challenges by leveraging the powerful latent space representation and reconstruction capabilities of Variational Autoencoders (VAE). However, applying VAE to UHD image restoration presents challenges: 1) High-performing VAEs have large parameter sizes, leading to significant carbon footprints; 2) The self-reconstruction property of VAE hinders bridging the domain gap between clean and degraded images; 3) Latent encoding in VAE can lose high-frequency information, compromising image detail. To overcome these challenges, we propose a frequency enhanced VAE UHD image restoration framework by integrating frequency priors. First, we design the Fourier-based lightweight frequency learning within the VAE to improve parameter efficiency. Then, we introduce a wavelet-based adapter that extracts multi-scale image information and employs frequency-aware adaptive modulation to bridge the domain gap by integrating degraded image data into the pre-trained VAE. Additionally, the adapter injects high-frequency information into the VAE decoder, enhancing detail in the restored images. In this way, our method effectively combines the powerful latent space representation with frequency priors to enhance UHD image restoration. Extensive experiments on various UHD image restoration tasks show that our method surpasses state-of-the-art methods both qualitatively and quantitatively.

Dynamic 3D interaction has been attracting a lot of attention recently. However, creating such 4D content remains challenging. One solution is to animate 3D scenes with physics-based simulation, which requires manually assigning precise physical properties to the object or the simulated results would become unnatural. Another solution is to learn the deformation of 3D objects with the distillation of video generative models, which, however, tends to produce 3D videos with small and discontinuous motions due to the inappropriate extraction and application of physics priors. In this work, to combine the strengths and complementing shortcomings of the above two solutions, we propose to learn the physical properties of a material field with video diffusion priors, and then utilize a physics-based Material-Point-Method (MPM) simulator to generate 4D content with realistic motions. In particular, we propose motion distillation sampling to emphasize video motion information during distillation. In addition, to facilitate the optimization, we further propose a KAN-based material field with frame boosting. Experimental results demonstrate that our method enjoys more realistic motions than state-of-the-arts do.

Diffusion models for garment-centric human generation from text or image prompts have garnered emerging attention for their great application potential. However, existing methods often face a dilemma: lightweight approaches, such as adapters, are prone to generate inconsistent textures; while finetune-based methods involve high training costs and struggle to maintain the generalization capabilities of pretrained diffusion models, limiting their performance across diverse scenarios. To address these challenges, we propose DreamFit, which incorporates a lightweight Anything-Dressing Encoder specifically tailored for the garment-centric human generation. DreamFit has three key advantages: (1) Lightweight training: with the proposed adaptive attention and LoRA modules, DreamFit significantly minimizes the model complexity to 83.4M trainable parameters. (2) Anything-Dressing: Our model generalizes surprisingly well to a wide range of (non-)garments, creative styles, and prompt instructions, consistently delivering high-quality results across diverse scenarios. (3) Plug-and-play: DreamFit is engineered for smooth integration with any community control plugins for diffusion models, ensuring easy compatibility and minimizing adoption barriers. To further enhance generation quality, DreamFit leverages pretrained large multi-modal models (LMMs) to enrich the prompt with fine-grained garment descriptions, thereby reducing the prompt gap between training and inference. We conduct comprehensive experiments on both 768 x 512 high-resolution benchmarks and in-the-wild images. DreamFit surpasses all existing methods, highlighting its state-of-the-art capabilities of garment-centric human generation.

Recent years have witnessed the remarkable success of Text-to-3D generation, particularly with the rise of mainstream conditional diffusion models (DMs). Though achieving substantial progress, existing methods still face a knotty "human preference" dilemma, that is the 3D contents generated by the models often deviate greatly from the desired effects (e.g., perspective, aesthetics, shading, appearance, etc.) due to the lack of attention to human preferences. To mitigate the limitation of data deficiency and enable human preference learning, we first elaborately curate the HP3D, a text-to-3D dataset with expert preference annotations which is initally captioned by the multimodal large model LLava and then refined by human expert. Based on such a brand-new HP3D, we further propose DreamAlign, a reward-free method that does not require designing any complex reward models whereas only by introducing a light-weight lora adapter and then designing a novel direct 3D preference optimization (D-3DPO) algorithm for training. Moreover, in the stage of text-to-3D we design an additional Preference Contrastive Feedback training for score distillation sampling, which enables the generated 3D objects to align the human preferences (e.g., aesthetics, material, etc.). Extensive experiments demonstrate that DreamAlign consistently achieves state-of-the-art performance on generative effects and human preference alignment across various benchmark evaluations.

Recent advancements in text-to-3D generation can generate neural radiance fields (NeRFs) with score distillation sampling, enabling 3D asset creation without real-world data capture. With the rapid advancement in NeRF generation quality, protecting the copyright of the generated NeRF has become increasingly important. While prior works can watermark NeRFs in a post-generation way, they suffer from two vulnerabilities. First, a delay lies between NeRF generation and watermarking because the secret message is embedded into the NeRF model post-generation through fine-tuning. Second, generating a non-watermarked NeRF as an intermediate creates a potential vulnerability for theft. To address both issues, we propose Dreamark to embed a secret message by backdooring the NeRF during NeRF generation. In detail, we first pre-train a watermark decoder. Then, Dreamark generates backdoored NeRFs in a way that the target secret message can be verified by the pre-trained watermark decoder on an arbitrary trigger viewport. We evaluate the generation quality and watermark robustness against image- and model-level attacks. Extensive experiments show that the watermarking process will not degrade the generation quality, and the watermark achieves 90+% accuracy among both image-level attacks (e.g., Gaussian noise) and model-level attacks (e.g., pruning attack).