Lens flares arise from light reflection and refraction within sensor arrays, whose diverse types include glow, veiling glare, reflective flare and so on. Existing methods are specialized for one specific type only, and overlook the simultaneous occurrence of multiple typed lens flares, which is common in the real-world, e.g. coexistence of glow and displacement reflections from the same light source. These co-occurring lens flares cannot be effectively resolved by the simple combination of individual flare removal methods, since these coexisting flares originates from the same light source and are generated simultaneously within the same sensor array, exhibit a complex interdependence rather than simple additive relation. To model this interdependent flares’ relationship, our Nighttime Lens Flare Formation model is the first attempt to learn the intrinsic physical relationship between flares on the imaging plane. Building on this physical model, we introduce a solution to this joint flare removal task named Self-supervised Generation-based Lens Flare Removal Network (SGLFR-Net), which is self-supervised without pre-training. Specifically, the nighttime glow is detangled in PSF Rendering Network(PSFR-Net) based on PSF Rendering Prior, while the reflective flare is modelled in Texture Prior Based Reflection Flare Removal Network (TPRR-Net). Empirical evaluations demonstrate the effectiveness of the proposed method in both joint and individual glare removal tasks.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement
PDF ↗Occupancy prediction plays a pivotal role in autonomous driving (AD) due to its capabilities of fine-grained 3D perception and general object recognition. However, existing methods often incur high computational costs, which conflict with AD's real-time demand. To this end, we redirect the focus from accuracy only to both accuracy and efficiency. By conducting a head-to-head comparison of existing methods, we find it challenging to balance accuracy and efficiency. We identify a core issue for this challenge: the strong coupling between geometry and semantics. Specifically, the predicted geometric structure (e.g., depth) guides the projection of 2D image features into 3D voxel space, which significantly affects feature discriminability and subsequent semantic learning. To address this issue, we focus on two key aspects: model design and learning strategies. 1) For model design, we propose a dual-branch network that disentangles the representation of geometry and semantics. The voxel branch utilizes a novel re-parameterized large-kernel 3D convolution to refine geometric structure efficiently, while the BEV branch employs temporal fusion and BEV encoding for efficient semantic learning. 2) For learning strategies, we propose to separate geometric learning from semantic learning by the mixup of ground-truth and predicted depths. Our method achieves 39.4% mIoU at 20 FPS on Occ3D-nuScenes, showcasing a state-of-the-art balance between accuracy and efficiency.
Current out-of-distribution (OOD) detection methods typically assume balanced in-distribution (ID) data, while most real-world data follow a long-tailed distribution. Previous approaches to long-tailed OOD detection often involve balancing the ID data by reducing the semantics of head classes. However, this reduction can severely affect the classification accuracy of ID data. The main challenge of this task lies in the severe lack of features for tail classes, leading to confusion with OOD data. To tackle this issue, we introduce a novel Prioritizing Attention to Tail (PATT) method using augmentation instead of reduction. Our main intuition involves using a mixture of von Mises-Fisher (vMF) distributions to model the ID data and a temperature scaling module to boost the confidence of ID data. This enables us to generate infinite contrastive pairs, implicitly enhancing the semantics of ID classes while promoting differentiation between ID and OOD data. To further strengthen the detection of OOD data without compromising the classification performance of ID data, we propose feature calibration during the inference phase. By extracting an attention weight from the training set that prioritizes the tail classes and reduces the confidence in OOD data, we improve the OOD detection capability. Extensive experiments verified that our method outperforms the current state-of-the-art methods on various benchmarks.
MagicMan: Generative Novel View Synthesis of Humans with 3D-Aware Diffusion and Iterative Refinement
PDF ↗Existing works in single-image human reconstruction suffer from weak generalizability due to insufficient training data or 3D inconsistencies for a lack of comprehensive multi-view knowledge. In this paper, we introduce MagicMan, a human-specific multi-view diffusion model to generate high-quality novel views from a single reference image. As its core, we leverage a pre-trained 2D diffusion model as the generative prior for generalizability, with the parametric SMPL-X model as the 3D body prior to promote 3D awareness. To maintain consistency while generating denser views for improved 3D human reconstruction, we introduce hybrid multi-view attention to facilitate efficient and thorough information interchange across views. Besides, we present a geometry-aware dual branch to perform concurrent generation in both RGB and normal domains, further enhancing consistency via geometry cues. Last but not least, to address ill-shaped issues arising from inaccurate SMPL-X estimation, we propose a novel iterative refinement strategy, which progressively optimizes SMPL-X accuracy while enhancing the quality and consistency of the generated multi-views. Extensive experimental results demonstrate that our method significantly outperforms existing approaches in both novel view synthesis and subsequent 3D human reconstruction tasks.
DiffCalib: Reformulating Monocular Camera Calibration as Diffusion-Based Dense Incident Map Generation
PDF ↗Monocular camera calibration is a key precondition for numerous 3D vision applications. Despite considerable advancements, existing methods often hinge on specific assumptions and struggle to generalize across varied real-world scenarios, and the performance is limited by insufficient training data. Recently, diffusion models trained on expansive datasets have been confirmed to maintain the capability to generate diverse, high-quality images. This success suggests a strong potential of the models to effectively understand varied visual information. In this work, we leverage the comprehensive visual knowledge embedded in pre-trained diffusion models to enable more robust and accurate monocular camera intrinsic estimation. Specifically, we reformulate the problem of estimating the four degrees of freedom (4-DoF) of camera intrinsic parameters as a dense incident map generation task. The map details the angle of incidence for each pixel in the RGB image, and its format aligns well with the paradigm of diffusion models. The camera intrinsic then can be derived from the incident map with a simple non-learning RANSAC algorithm during inference. Moreover, to further enhance the performance, we jointly estimate a depth map to provide extra geometric information for the incident map estimation. Extensive experiments on multiple testing datasets demonstrates that our model achieves state-of-the-art performance, gaining up to a 40% reduction in prediction errors. Besides, the experiments also show that the precise camera intrinsic and depth maps estimated by our pipeline can greatly benefit practical applications such as 3D reconstruction from a single in-the-wild image.
In time-lapse microscopy, inherent noise significantly limits imaging sensitivity and increases measurement uncertainty. Due to the scarcity of clean data, zero-shot approaches have emerged as highly data-efficient solutions for microscopy denoising. However, existing methods typically process video frames independently, resulting in long training times and issues such as temporal noise and over-smoothing. In this paper, we introduce MDSR-Zero, a zero-shot online learning method designed for plug-and-play noise suppression and super-resolution of microscopy videos. Our approach leverages an efficient online training strategy that reuses denoising models from previous frames. By treating the video as a continuous stream, our model significantly reduces training time and ensures temporally consistent denoising. Additionally, we propose a novel loss function tailored for denoising in the context of super-resolution, which enhances the detail in the denoised results. Extensive experiments on both synthetic and real-world noise demonstrate that our method achieves state-of-the-art performance among zero-shot denoising approaches and is competitive with self-supervised methods. Notably, our method can reduce training time by up to 10x compared to the previous SOTA method.
Transformers have revolutionized the point cloud learning task, but the quadratic complexity hinders its extension to long sequence and makes a burden on limited computational resources. The recent advent of RWKV, a fresh breed of deep sequence models, has shown immense potential for sequence modeling in NLP tasks. In this paper, we present PointRWKV, a model of linear complexity derived from the RWKV model in the NLP field with necessary modifications for point cloud learning tasks. Specifically, taking the embedded point patches as input, we first propose to explore the global processing capabilities within PointRWKV blocks using modified multi-headed matrix-valued states and a dynamic attention recurrence mechanism. To extract local geometric features simultaneously, we design a parallel branch to encode the point cloud efficiently in a fixed radius near-neighbors graph with a graph stabilizer. Furthermore, we design PointRWKV as a multi-scale framework for hierarchical feature learning of 3D point clouds, facilitating various downstream tasks. Extensive experiments on different point cloud learning tasks show our proposed PointRWKV outperforms the transformer- and mamba-based counterparts, while significantly saving about 42\% FLOPs, demonstrating the potential option for constructing foundational 3D models.
Concept Bottleneck Models (CBMs) offer inherent interpretability by initially translating images into human-comprehensible concepts, followed by a linear combination of these concepts for classification. However, the annotation of concepts for visual recognition tasks requires extensive expert knowledge and labor, constraining the broad adoption of CBMs. Recent approaches have leveraged the knowledge of large language models to construct concept bottlenecks, with multimodal models like CLIP subsequently mapping image features into the concept feature space for classification. Despite this, the concepts produced by language models can be verbose and may introduce non-visual attributes, which hurts accuracy and interpretability. In this study, we investigate to avoid these issues by constructing CBMs directly from multimodal models. To this end, we adopt common words as base concept vocabulary and leverage auxiliary unlabeled images to construct a Vision-to-Concept (V2C) tokenizer that can explicitly quantize images into their most relevant visual concepts, thus creating a vision-oriented concept bottleneck tightly coupled with the multimodal model. This leads to our V2C-CBM which is training efficient and interpretable with high accuracy. Our V2C-CBM has matched or outperformed LLM-supervised CBMs on various visual classification benchmarks, validating the efficacy of our approach.
The rapid progress of multimedia technology has led to an increased focus on enhancing the quality of experience (QoE) for video. Specifically, the demand for low-latency and high-quality decoding has grown significantly. Compressed Video Quality Enhancement (CVQE) methods based on Deep Neural Networks (DNNs) have achieved remarkable success. However, most of the methods suffer from high computational complexity, thereby limiting their practicality in low-latency scenarios. Recently, Look-Up Table (LUT) methods have shown great efficiency, which makes them considerably promising in the field of low-latency CVQE. In this paper, we propose an efficient multi-frame deformable Look-Up Table structure for CVQE. Firstly, we design an efficient CNN to explore the inter-frame correlation and then predict the multi-scale convolution offsets. Secondly, we introduce a temporal feature extraction module and a multi-scale fusion module. We first exploit the predicted offsets to guide sampling for precise temporal alignment and extract multi-frame information. Then, higher quality frames are reconstructed from the fused multi-scale features. During the inference, we convert these two modules into LUTs to achieve a sound trade-off between model performance and computational complexity. Experiments demonstrate that our proposed method dramatically outperforms the state-of-the-art LUT-based methods, and obtains competitive performance compared to CNN-based methods with the capability to run in real-time(30fps) at 1080p resolution.
While recent works have achieved great success on one-shot 3D common object generation, high quality and fidelity 3D head generation from a single image remains a great challenge. Previous text-based methods for generating 3D heads were limited by text descriptions and image-based methods struggled to produce high-quality head geometry. To handle this challenging problem, we propose a novel framework, ID-Sculpt, to generate high-quality 3D heads while preserving their identities. Our work incorporates the identity information of the portrait image into three parts: 1) geometry initialization, 2) geometry sculpting, and 3) texture generation stages. Given a reference portrait image, we first align the identity features with text features to realize ID-aware guidance enhancement, which contains the control signals representing the face information. We then use the canny map, ID features of the portrait image, and a pre-trained text-to-normal/depth diffusion model to generate ID-aware geometry supervision and 3D-GAN inversion is employed to generate ID-aware geometry initialization. Furthermore, with the ability to inject identity information into 3D head generation, we use ID-aware guidance to calculate ID-aware Score Distillation (ISD) for geometry sculpting. For texture generation, we adopt the ID Consistent Texture Inpainting and Refinement which progressively expands the view for texture inpainting to obtain an initialization UV texture map. We then use the id-aware guidance to provide image-level supervision for noisy multi-view images to obtain a refined texture map. Extensive experiments demonstrate that we can generate high-quality 3D heads with accurate geometry and texture from a single in-the-wild portrait image.
Image inpainting is an important image generation task, which aims to restore corrupted image from partial visible area. Recently, diffusion Schrödinger bridge methods effectively tackle this task by modeling the translation between corrupted and target images as a diffusion Schrödinger bridge process along a noising schedule path. Although these methods have shown superior performance, in this paper, we find that 1) existing methods suffer from a schedule-restoration mismatching issue, i.e., the theoretical schedule and practical restoration processes usually exist a large discrepancy, which theoretically results in the schedule not fully leveraged for restoring images; and 2) the key reason causing such issue is that the restoration process of all pixels are actually asynchronous but existing methods set a synchronous noise schedule to them, i.e., all pixels shares the same noise schedule. To this end, we propose a schedule-Asynchronous Diffusion Schrödinger Bridge (AsyncDSB) for image inpainting. Our insight is preferentially scheduling pixels with high frequency (i.e., large gradients) and then low frequency (i.e., small gradients). Based on this insight, given a corrupted image, we first train a network to predict its gradient map in corrupted area. Then, we regard the predicted image gradient as prior and design a simple yet effective pixel-asynchronous noise schedule strategy to enhance the diffusion Schrödinger bridge. Thanks to the asynchronous schedule at pixels, the temporal interdependence of restoration process between pixels can be fully characterized for high-quality image inpainting. Experiments on real-world datasets show that our AsyncDSB achieves superior performance, especially on FID with around 3% ∼ 14% improvement over state-of-the-art baseline methods.
CoDTS: Enhancing Sparsely Supervised Collaborative Perception with a Dual Teacher-Student Framework
PDF ↗Current collaborative perception methods often rely on fully annotated datasets, which can be expensive to obtain in practical situations. To reduce annotation costs, some works adopt sparsely supervised learning techniques and generate pseudo labels for the missing instances. However, these methods fail to achieve an optimal confidence threshold that harmonizes the quality and quantity of pseudo labels. To address this issue, we propose an end-to-end Collaborative perception Dual Teacher-Student framework (CoDTS), which employs adaptive complementary learning to produce both high-quality and high-quantity pseudo labels. Specifically, the Main Foreground Mining (MFM) module generates high-quality pseudo labels based on the prediction of the static teacher. Subsequently, the Supplement Foreground Mining (SFM) module ensures a balance between the quality and quantity of pseudo labels by adaptively identifying missing instances based on the prediction of the dynamic teacher. Additionally, the Neighbor Anchor Sampling (NAS) module is incorporated to enhance the representation of pseudo labels. To promote the adaptive complementary learning, we implement a staged training strategy that trains the student and dynamic teacher in a mutually beneficial manner. Extensive experiments demonstrate that the CoDTS effectively ensures an optimal balance of pseudo labels in both quality and quantity, establishing a new state-of-the-art in sparsely supervised collaborative perception.
The recent Segment Anything Model (SAM) has emerged as a new paradigmatic vision foundation model, showcasing potent zero-shot generalization and flexible prompting. Despite SAM finding applications and adaptations in various domains, its primary limitation lies in the inability to grasp object semantics. In this paper, we present Sambor to seamlessly integrate SAM with the open-vocabulary object detector in an end-to-end framework. While retaining all the remarkable capabilities inherent to SAM, we boost it to detect arbitrary objects from human inputs like category names or reference expressions. Building upon the SAM image encoder, we introduce a novel SideFormer module designed to acquire SAM features adept at perceiving objects and inject comprehensive semantic information for recognition. In addition, we devise an Open-set RPN that leverages SAM proposals to assist in finding potential objects. Consequently, Sambor enables the open-vocabulary detector to equally focus on generalizing both localization and classification sub-tasks. Our approach demonstrates superior zero-shot performance across benchmarks, including COCO and LVIS, proving highly competitive against previous state-of-the-art methods. We aspire for this work to serve as a meaningful endeavor in endowing SAM to recognize diverse object categories and advancing open-vocabulary learning with the support of vision foundation models.
There are two crucial aspects of reliable autonomous driving systems: the reasoning behind decision-making and the precision of environmental perception. This paper introduces DME-Driver, a new autonomous driving system that enhances performance and robustness by fully leveraging the two crucial aspects. This system comprises two main models. The first, the Decision Maker, is responsible for providing logical driving instructions. The second, the Executor, receives these instructions and generates precise control signals for the vehicles. To ensure explainable and reliable driving decisions, we build the Decision-Maker based on a large vision language model. This model follows the logic employed by experienced human drivers and simulates making decisions in a safe and reasonable manner. On the other hand, the generation of accurate control signals relies on precise and detailed environmental perception, where 3D scene perception models excel. Therefore, a planning-oriented perception model is employed as the Executor. It translates the logical decisions made by the Decision-Maker into accurate control signals for the self-driving cars. To effectively train the proposed system, a new dataset named Human-driver Behavior and Decision-making (HBD) dataset has been collected. This dataset encompasses a diverse range of human driver behaviors and their underlying motivations. By leveraging this dataset, our system achieves high-precision planning accuracy through a logical thinking process.
ProPose: Probabilistic 3D Human Pose Estimation with Instance-Level Distribution and Normalizing Flow
PDF ↗3D Human Pose Estimation (HPE) is a one-to-many problem by nature, making it challenging to estimate an accurate 3D pose from a single 2D pose. Some prior works have attempted to tackle this problem by using a conditional generative network. They generate 3D poses from a given 2D pose with noises from a standard Gaussian distribution, while the depth distribution is dependent on each posture and more complex than the standard Gaussian distribution. This may lead to inaccurate distribution learning. In this paper, we propose a probabilistic framework called ProPose to address this issue. ProPose employs Pose Instance-Level Gaussian Distribution (PILGD) derived from 3D pose-based self-representation learning to obtain reliable distribution which is able to address pose-dependent depth distribution. To access this PILGD, we utilize normalizing flow, which learns a mapping function between the PILGD and a 2D Pose-Adaptive Gaussian Distribution (PAGD). This converts the problem of directly estimating 3D poses from 2D poses to a mapping problem between PILGD and PAGD using a normalizing flow. Extensive experiments show the advantages of utilizing the PILGD and PAGD. ProPose achieves comparable performances to previous state-of-the-art probabilistic methods in a multi-hypothesis setting. Notably, ProPose in a single-hypothesis setting demonstrates comparable generalization ability to existing state-of-the-art deterministic methods.
We propose GGS, a Generalizable Gaussian Splatting method for Autonomous Driving that can achieve realistic rendering under large viewpoint changes. Previous generalizable 3D gaussian splatting methods are limited to rendering novel views that are very close to the original pair of images, which cannot handle large difference in viewpoint. Especially in autonomous driving scenarios, images are typically collected from a single lane. The limited training perspective makes rendering images of a different lane very challenging. To further improve the rendering capability of GGS under large viewpoint changes, we introduce a novel virtual lane generation module into GSS method to enable high-quality lane switching even without a multi-lane dataset. Besides, we design a diffusion loss to supervise the generation of virtual lane images to further address the problem of data lacking in the virtual lanes. Finally, we also propose a depth refinement module to optimize depth estimation in the GSS model. Extensive validation of our method, compared to existing approaches, demonstrates state-of-the-art performance.
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.
LLaVA Needs More Knowledge: Retrieval Augmented Natural Language Generation with Knowledge Graph for Explaining Thoracic Pathologies
PDF ↗Generating Natural Language Explanations (NLEs) for model predictions on medical images, particularly those depicting thoracic pathologies, remains a critical and challenging task. Existing methodologies often struggle due to general models' insufficient domain-specific medical knowledge and privacy concerns associated with retrieval-based augmentation techniques. To address these issues, we propose a novel Vision-Language framework augmented with a Knowledge Graph (KG)-based datastore, which enhances the model's understanding by incorporating additional domain-specific medical knowledge essential for generating accurate and informative NLEs. Our framework employs a KG-based retrieval mechanism that not only improves the precision of the generated explanations but also preserves data privacy by avoiding direct data retrieval. The KG datastore is designed as a plug-and-play module, allowing for seamless integration with various model architectures. We introduce and evaluate three distinct frameworks within this paradigm: KG-LLaVA, which integrates the pre-trained LLaVA model with KG-RAG; Med-XPT, a custom framework combining MedCLIP, a transformer-based projector, and GPT-2; and Bio-LLaVA, which adapts LLaVA by incorporating the Bio-ViT-L vision model. These frameworks are validated on the MIMIC-NLE dataset, where they achieve state-of-the-art results, underscoring the effectiveness of KG augmentation in generating high-quality NLEs for thoracic pathologies.
VTG-LLM: Integrating Timestamp Knowledge into Video LLMs for Enhanced Video Temporal Grounding
PDF ↗Video Temporal Grounding (VTG) strives to accurately pinpoint event timestamps in a specific video using linguistic queries, significantly impacting downstream tasks like video browsing and editing. Unlike traditional task-specific models, Video Large Language Models (video LLMs) can handle multiple tasks concurrently in a zero-shot manner. Consequently, exploring the application of video LLMs for VTG tasks has become a burgeoning research area. However, despite considerable advancements in video content understanding, video LLMs often struggle to accurately pinpoint timestamps within videos, limiting their effectiveness in VTG tasks. To address this, we introduce VTG-LLM, a model designed to enhance video LLMs' timestamp localization abilities. Our approach includes: (1) effectively integrating timestamp knowledge into visual tokens; (2) incorporating absolute-time tokens to manage timestamp knowledge without concept shifts; and (3) introducing a lightweight, high-performance, slot-based token compression technique designed to accommodate the demands of a large number of frames to be sampled for VTG tasks. Additionally, we present VTG-IT-120K, a collection of publicly available VTG datasets that we have re-annotated to improve upon low-quality annotations. Our comprehensive experiments demonstrate the superior performance of VTG-LLM in comparison to other video LLM methods across a variety of VTG tasks.
3D Gaussian Splatting (3DGS) has been proven to exhibit exceptional performance in reconstructing 3D scenes. However, the effectiveness of 3DGS heavily relies on sharp images, and fulfilling this requirement presents challenges in real-world scenarios particularly when utilizing fast-moving cameras. This limitation severely constrains the practical application of 3DGS and may compromise the feasibility of real-time reconstruction. To mitigate these challenges, we proposed Spike Gaussian Splatting (SpikeGS), the first framework that integrates the Bayer-pattern spike streams into the 3DGS pipeline to reconstruct 3D scenes captured by a fast-moving high temporal color spike camera in one second. With accumulation rasterization, interval supervision, and a special designed pipeline, SpikeGS realizes continuous spatiotemporal perception while extracts detailed structure and texture from Bayer-pattern spike stream which is unstable and lacks details. Extensive experiments on both synthetic and real-world datasets demonstrate the superiority of SpikeGS compared with existing spike-based and deblur 3D scene reconstruction methods.