论文检索

Yu Bing Luo, Nian Shi, Jia Qin, Zekai Ji, Pinle Qin, Jianchao Zeng, Jianghui Cai

Downsampling is essential in semantic segmentation for reducing computational cost and guiding the learning of class-discriminative features. Existing models typically rely on strided convolutions or patch splitting to obtain features with lower resolution. However, we observe that such operations often introduce edge jagging and texture degradation, the underlying cause is that aliasing of the high frequency induces phase distortion. We conducted a systematic analysis of phase distortion and identified two key properties: spatial non-uniformity (concentrated near boundaries) and directional sparsity (accumulated along a few dominant directions). These properties cause crucial high-frequency cues to be misrepresented or lost during sampling. To address this issue, we propose a frequency aware filter consisting of two complementary modules: a dynamic Gaussian kernel (DGK) and a learnable Gabor-based frequency selector (LFS). To mitigate spatial non-uniformity, the DGK predicts edge normals from gradients, applies strong low-pass filtering along the normal direction, and leaves the tangential direction virtually untouched, thereby suppressing phase distortion while preserving contour continuity. To handle directional sparsity, the Learnable Gabor Selector (LFS) then performs directional band-pass filtering to attenuate residual aliasing peaks and adaptively boost informative texture. We further introduce phase-error energy (PE) to quantify distortion severity. Visualization and quantitative results demonstrate that frequency-aware filter offers a plug-and-play remedy for aliasing, yielding sharper boundaries and consistent gains across datasets.

Zhiguo Lu, Jianwen Lou, Mingjun Ma, Hairong Jin, Youyi Zheng, Kun Zhou

3D teeth segmentation, involving the localization of tooth instances and their semantic categorization in 3D dental models, is a critical yet challenging task in digital dentistry due to the complexity of real-world dentition. In this paper, we propose 3DTeethSAM, an adaptation of the Segment Anything Model 2 (SAM2) for 3D teeth segmentation. SAM2 is a pretrained foundation model for image and video segmentation, demonstrating a strong backbone in various downstream scenarios. To adapt SAM2 for 3D teeth data, we render images of 3D teeth models from predefined views, apply SAM2 for 2D segmentation, and reconstruct 3D results using 2D-3D projections. Since SAM2's performance depends on input prompts and its initial outputs often have deficiencies, and given its class-agnostic nature, we introduce three light-weight learnable modules: (1) a prompt embedding generator to derive prompt embeddings from image embeddings for accurate mask decoding, (2) a mask refiner to enhance SAM2's initial segmentation results, and (3) a mask classifier to categorize the generated masks. Additionally, we incorporate Deformable Global Attention Plugins (DGAP) into SAM2's image encoder. The DGAP enhances both the segmentation accuracy and the speed of the training process. Our method has been validated on the 3DTeethSeg benchmark, achieving an IoU of 91.90% on high-resolution 3D teeth meshes, establishing a new state-of-the-art in the field.

Wei Lu, Xue Yang, Si-Bao Chen

Light-weight neural networks for remote sensing (RS) visual analysis must overcome two inherent redundancies: spatial redundancy from vast, homogeneous backgrounds, and channel redundancy, where extreme scale variations render a single feature space inefficient. Existing models, often designed for natural images, fail to address this dual challenge in RS scenarios. To bridge this gap, we propose LWGANet, a light-weight backbone engineered for RS-specific properties. LWGANet introduces two core innovations: a Top-K Global Feature Interaction (TGFI) module that mitigates spatial redundancy by focusing computation on salient regions, and a Light-Weight Grouped Attention (LWGA) module that resolves channel redundancy by partitioning channels into specialized, scale-specific pathways. By synergistically resolving these core inefficiencies, LWGANet achieves a superior trade-off between feature representation quality and computational cost. Extensive experiments on twelve diverse datasets across four major RS tasks—scene classification, oriented object detection, semantic segmentation, and change detection—demonstrate that LWGANet consistently outperforms state-of-the-art light-weight backbones in both accuracy and efficiency. Our work establishes a new, robust baseline for efficient visual analysis in RS images.

Haocheng Lu, Nan Zhang, Wei Tao, Xiaoyang Qu, Guokuan Li, Jiguang Wan, Jianzong Wang

Streaming video question answering (Streaming Video QA) poses distinct challenges for multimodal large language models (MLLMs), as video frames arrive sequentially and user queries can be issued at arbitrary timepoints. Existing solutions relying on fixed-size memory or naive compression often suffer from context loss or memory overflow, limiting their effectiveness in long-form, real-time scenarios.We present Vista, a novel framework for scene-aware streaming video QA that enables efficient and scalable reasoning over continuous video streams. The innovation of Vista can be summarized in three aspects: (1) Scene-aware segmentation. Vista dynamically clusters incoming frames into temporally and visually coherent scene units. (2) Scene-aware compression. Each scene is compressed into a compact token representation and stored in GPU memory for efficient index-based retrieval, while the full-resolution frames are offloaded to CPU memory. (3) Scene-aware recall. Upon receiving a question, relevant scenes are selectively recalled and reintegrated into the model’s input space, enabling both efficiency and completeness. Vista is model-agnostic and integrates seamlessly with a variety of vision-language backbones, enabling long-context reasoning without compromising latency or memory efficiency. Extensive experiments on StreamingBench demonstrate that Vista achieves state-of-the-art performance, establishing a strong baseline for real-world streaming video understanding.

Yanzhen Liu, Sutuke Yibulayimu, Yang Zhou, Yudi Sang, Yu Wang

Fracture injuries often lead to complex bone fragmentations, posing significant challenges for accurate segmentation in surgical planning and trauma assessment. Manual annotation of each fragment is time-consuming and inconsistent, while existing automated methods often fail to separate individual fragments due to the wide variation in fracture types, irregular fracture surface, and close inter-fragment contact. To address these challenges, we introduce FracSegmentator, a deep learning approach for bone fragment instance segmentation. The model takes extracted bone regions in CT as input and isolates individual fragments by identifying fracture surfaces and separating closely contacting structures. Central to our approach is a Trauma-Prior-Guided Contrastive Learning module, which incorporates clinical knowledge through memory-based attention to better distinguish fractured surfaces from healthy regions. We evaluate FracSegmentator on four datasets that cover a range of anatomical sites and fracture patterns. The method achieves state-of-the-art results across all datasets and demonstrates strong generalization capabilities. By delivering accurate and efficient fragment-level segmentation, FracSegmentator supports critical downstream tasks such as automated fracture diagnosis, surgical planning, and preoperative reduction simulation.

Xuexun Liu, Xiaoxu Xu, Qiudan Zhang, Lin Ma, Xu Wang

Weakly supervised 3D instance segmentation is essential for 3D scene understanding, especially as the growing scale of data and high annotation costs associated with fully supervised approaches. Existing methods primarily rely on two forms of weak supervision: one-thing-one-click annotations and bounding box annotations, both of which aim to reduce labeling efforts. However, these approaches still encounter limitations, including labor-intensive annotation processes, high complexity, and reliance on expert annotators. To address these challenges, we propose DBGroup, a two-stage weakly supervised 3D instance segmentation framework that leverages scene-level annotations as a more efficient and scalable alternative. In the first stage, we introduce a Dual-Branch Point Grouping module to generate pseudo labels guided by semantic and mask cues extracted from multi-view images. To further improve label quality, we develop two refinement strategies: Granularity-Aware Instance Merging and Semantic Selection and Propagation. The second stage involves multi-round self-training on an end-to-end instance segmentation network using the refined pseudo-labels. Additionally, we introduce an Instance Mask Filter strategy to address inconsistencies within the pseudo labels. Extensive experiments demonstrate that DBGroup achieves competitive performance compared to sparse-point-level supervised 3D instance segmentation methods, while surpassing state-of-the-art scene-level supervised 3D semantic segmentation approaches.

Shang Liu, Chenjie Cao, Chaohui Yu, Wen Qian, Jing Wang, Fan Wang

Despite the remarkable developments achieved by recent 3D generation works, scaling these methods to geographic extents, such as modeling thousands of square kilometers of Earth’s surface, remains an open challenge. We address this through a dual innovation in data infrastructure and model architecture. First, we introduce Aerial-Earth3D, the largest 3D aerial dataset to date, consisting of 50k curated scenes (each measuring 600m) captured across the U.S. mainland, comprising 45M multi-view Google Earth frames. Each scene provides pose-annotated multi-view images, depth maps, normals, semantic segmentation, and camera poses, with explicit quality control to ensure terrain diversity. Building on this foundation, we propose EarthCrafter, a tailored framework for large-scale 3D Earth generation via sparse-decoupled latent diffusion. Our architecture separates structural and textural generation: 1) Dual sparse 3D-VAEs compress high-resolution geometric voxels and textural 2D Gaussian Splats (2DGS) into compact latent spaces, largely alleviating the costly computation suffering from vast geographic scales while preserving critical information. 2) We propose condition-aware flow matching models trained on mixed inputs (semantics, images, or neither) to flexibly model latent geometry and texture features independently. Extensive experiments demonstrate that EarthCrafter performs substantially better in extremely large-scale generation. The framework further supports versatile applications, from semantic-guided urban layout generation to unconditional terrain synthesis, while maintaining geographic plausibility through our rich data priors from Aerial-Earth3D.

Lin Liu, Huixia Ben, Shuo Wang, Jinda Lu, Junxiang Qiu, Shengeng Tang, Yanbin Hao

Controllable generative models have been widely used to improve the realism of synthetic visual content. However, such models must handle control conditions and content generation computational requirements, resulting in generally low generation efficiency. To address this issue, we propose a Hybrid-Grained Cache (HGC) approach that reduces computational overhead by adopting cache strategies with different granularities at different computational stages. Specifically, (1) we use a coarse-grained cache (block-level) based on feature reuse to dynamically bypass redundant computations in encoder-decoder blocks between each step of model reasoning. (2) We design a fine-grained cache (prompt-level) that acts within a module, where the fine-grained cache reuses cross-attention maps within consecutive reasoning steps and extends them to the corresponding module computations of adjacent steps. These caches of different granularities can be seamlessly integrated into each computational link of the controllable generation process. We verify the effectiveness of HGC on four benchmark datasets, especially its advantages in balancing generation efficiency and visual quality. For example, on the COCO-Stuff segmentation benchmark, our HGC significantly reduces the computational cost (MACs) by 63% (from 18.22T → 6.70T↓), while keeping the loss of semantic fidelity (quantized performance degradation) within 1.5%.

Delong Liu, Haotian Hou, Zhaohui Hou, Shihao Han, Zhiyuan Huang, Mingjie Zhan, Fei Su, Zhicheng Zhao

Precise and controllable image editing, especially object removal and insertion, represents one of the most common demands in image manipulation. However, existing methods suffer from severe limitations. Mask-based inpainting often introduces visual artifacts and semantic inconsistencies, while instruction-based approaches lack accurate spatial control and tend to unintentionally modify background regions. To address these issues, we propose two key contributions. First, we develop a fully automated and self-improving pipeline for synthetic data generation. This pipeline utilizes a Large Language Model (LLM) to generate diverse prompts, a Diffusion Transformer (DiT) fine-tuned evolutionarily to synthesize high-quality images, and a Multimodal LLM (MLLM) combined with open-set object detector for automated quality control and annotation. This process produces the Remove/Add Dataset (RAD), consisting of over 514,510 high-quality image pairs, each richly annotated with bounding boxes, segmentation masks, and a variety of editing instructions. Second, based on RAD, we introduce Remove/Add Anything (RAA), a novel editing framework with precise spatial control. Built upon a diffusion-based inpainting model, RAA achieves high editing accuracy by conditioning on both textual instructions and an explicitly defined region of interest (ROI), enabling efficient fine-tuning while maintaining global visual coherence. Extensive experiments demonstrate that RAA significantly outperforms existing open-source methods on both addition and removal tasks, and even slightly surpasses costly proprietary models.

Zheng Lin, Nan Zhou, Yuhan Wang, Bojian Zhang

Interactive segmentation aims to delineate a user-specified target in an image by leveraging positive and negative clicks. While effective on natural images, existing methods often fail in remote sensing scenarios, where satellite imagery is characterized by ultra-high resolution, sparse object distribution, and significant scale variation. These factors hinder accurate segmentation of fine-grained targets like roads, buildings, and aircraft. To overcome these problems, we propose CrossCut, a novel interactive segmentation framework tailored for remote sensing imagery. Unlike previous approaches that either process the entire image or treat each patch independently, CrossCut enables simultaneous segmentation across multiple patches by propagating user click information to all patches. This design allows the model to fully utilize click guidance regardless of object location, effectively resolving the challenge of inter-patch information isolation. Furthermore, CrossCut supports flexible inference by allowing segmentation results from different patch configurations to be fused, enhancing both accuracy and robustness. Extensive evaluations across multiple remote sensing datasets demonstrate that CrossCut achieves state-of-the-art performance. Quantitative results and visualizations show that CrossCut significantly advances the field of interactive segmentation for remote sensing imagery.

Jianghang Lin, Yilin Lu, Chaoyang Zhu, Yunhang Shen, Shengchuan Zhang, Liujuan Cao

Semi-Supervised Instance Segmentation (SSIS) involves classifying and grouping image pixels into distinct object instances using limited labeled data alongside large-scale unlabeled data. A major challenge in SSIS lies in the inherent noise of pseudo-labels, particularly when class and mask qualities are coupled into a single confidence score for filtering. Such coupling often results in sub-optimal trade-offs between semantic accuracy and spatial precision. To address this, we propose a novel Pseudo-Label Decoupling and Correction (PL-DC) framework, which explicitly decouples and enhances the pseudo-label selection process for SSIS. At the instance level, we introduce a Decoupled Filtering with Adaptive Class-Aware Thresholds mechanism, which independently evaluates class and mask qualities using category-specific thresholds updated via exponential moving averages. At the category level, we design a Dynamic Instance Category Correction module that reassigns ambiguous class pseudo-label by leveraging semantic prototypes and consistency alignment. At the pixel level, a Pixel-Level Mask Uncertainty-Aware mechanism is applied to suppress the influence of unreliable pixels during mask supervision, further improving the robustness against pixel-wise noise. Extensive experiments on COCO and Cityscapes datasets demonstrate that the proposed PL-DC achieves significant performance improvements, setting new state-of-the-art results. Notably, PL-DC achieves gains of +11.7 mAP with just 1% labeled COCO data and +16.4 mAP with 5% Cityscapes labels, showing its effectiveness under extremely low-label regimes.

Jia Lin, Xiaofei Zhou, Jiyuan Liu, Runmin Cong, Guodao Zhang, Zhi Liu, Jiyong Zhang

Recently segment anything model (SAM) has attracted widespread concerns, and it is often treated as a vision foundation model for universal segmentation. Some researchers have attempted to directly apply the foundation model to the RGB-D video salient object detection (RGB-D VSOD) task, which often encounters three challenges, including the dependence on manual prompts, the high memory consumption of sequential adapters, and the computational burden of memory attention. To address the limitations, we propose a novel method, namely Segment Anything Model with Depth-guided Adaptive Queries (SAM-DAQ), which adapts SAM2 to pop-out salient objects from videos by seamlessly integrating depth and temporal cues within a unified framework. Firstly, we deploy a parallel adapter-based multi-modal image encoder (PAMIE), which incorporates several depth-guided parallel adapters (DPAs) in a skip-connection way. Remarkably, we fine-tune the frozen SAM encoder under prompt-free conditions, where the DPA utilizes depth cues to facilitate the fusion of multi-modal features. Secondly, we deploy a query-driven temporal memory (QTM) module, which unifies the memory bank and prompt embeddings into a learnable pipeline. Concretely, by leveraging both frame-level queries and video-level queries simultaneously, the QTM module can not only selectively extract temporal consistency features but also iteratively update the temporal representations of the queries. Extensive experiments are conducted on three RGB-D VSOD datasets, and the results show that the proposed SAM-DAQ consistently outperforms state-of-the-art methods in terms of all evaluation metrics.

Yun Liang, ZiHao Wu, Run Zheng, Shuai Xie, Bo Hong, Yishen Lin

Floor plan recognition requires accurate segmentation and classification of entrance doors, outer contours (walls and windows) and inner contours (various room types) , despite strong spatial dependencies and large stylistic differences between different datasets. To overcome these challenges, we propose FloorPlanFormer, a multi-task learning network divided into three phases: the first phase introduces a Swin Transformer backbone with a pixel decoder to extract fine-grained pixel-level semantics; the second phase employs prompt encoder and mask decoder, and a novel Global Contextual Attention Module (GCAM) is designed to generate clear, high-quality outer contour masks; the third stage uses mask transformer decoder to recognize targets and designs a Masked Feature Refinement Module (MFRM) to accurately delineate the inner contour by modeling the relationship between the local inner and outer contours. Finally, we constructed FloorPlan8K, a dataset containing 8200 images and 77434 instances, on which our model was trained and evaluated, and the results greatly outperformed the state-of-the-art general segmentation methods and specialized methods.

Shiyi Liang, Xinyuan Chang, Changjie Wu, Huiyuan Yan, Yifan Bai, Xinran Liu, Hang Zhang, Yujian Yuan, Shuang Zeng, Mu Xu 等

Safe autonomous driving requires both accurate HD map construction and persistent awareness of traffic rules, even when their associated signs are no longer visible. However, existing methods either focus solely on geometric elements or treat rules as temporary classifications, failing to capture their persistent effectiveness across extended driving sequences. In this paper, we present PAMR (Persistent Autoregressive Mapping with Traffic Rules), a novel framework that performs autoregressive co-construction of lane vectors and traffic rules from visual observations. Our approach introduces two key mechanisms: Map-Rule Co-Construction for processing driving scenes in temporal segments, and Map-Rule Cache for maintaining rule consistency across these segments. To properly evaluate continuous and consistent map generation, we develop MapDRv2, featuring improved lane geometry annotations. Extensive experiments demonstrate that PAMR achieves superior performance in joint vector-rule mapping tasks, while maintaining persistent rule effectiveness throughout extended driving sequences.

Qifan Liang, Junlin Li, Zhen Han, Xihao Wang, Zhongyuan Wang, Bin Mei

Electrocautery or lasers will inevitably generate surgical smoke, which hinders the visual guidance of laparoscopic videos for surgical procedures. The surgical smoke can be classified into different types based on its motion patterns, leading to distinctive spatio-temporal characteristics across smoky laparoscopic videos. However, existing desmoking methods fail to account for such smoke-type-specific distinctions. Therefore, we propose the first Smoke-Type-Aware Laparoscopic Video Desmoking Network (STANet) by introducing two smoke types: Diffusion Smoke and Ambient Smoke. Specifically, a smoke mask segmentation sub-network is designed to jointly conduct smoke mask and smoke type predictions based on the attention-weighted mask aggregation, while a smokeless video reconstruction sub-network is proposed to perform specially desmoking on smoky features guided by two types of smoke mask. To address the entanglement challenges of two smoke types, we further embed a coarse-to-fine disentanglement module into the mask segmentation sub-network, which yields more accurate disentangled masks through the smoke-type-aware cross attention between non-entangled and entangled regions. In addition, we also construct the first large-scale synthetic video desmoking dataset with smoke type annotations. Extensive experiments demonstrate that our method not only outperforms state-of-the-art approaches in quality evaluations, but also exhibits superior generalization across multiple downstream surgical tasks.

Jiawei Lian, Zhengxue Wang, Wentao Qu, Haobo Jiang, Le Hui, Jian Yang

Point cloud semantic segmentation is fundamental to 3D scene understanding, but dense annotation requirements limit scalability. Although recent label propagation and contrastive learning methods enhance local consistency, the incomplete object coverage caused by sparse annotations hinders global context modeling, ultimately limiting overall performance. To this end, we propose a diffusion-based contextual reconstruction framework for point cloud semantic segmentation with limited annotations. At its core, our framework guides denoising with semantic predictions, using better context reconstruction to enhance the conditional model for better segmentation. Specifically, our contributions include: (1) Diffusion-based segmentation framework: reconstructs contextual semantics from noise under conditional guidance, sharing the decoder with the segmentation module for robust contextual semantic learning. (2) Dynamically aggregates local context from segmentation features and guides denoising with global spatial structure, significantly enhancing denoising quality and contextual awareness. Notably, we pioneer diffusion models for 3D semantic segmentation with limited annotations, enabling efficient single-step inference. Experiments show robustness across varying annotation ratios and state-of-the-art performance on benchmarks.

Zhenghua Li, Hang Chen, Zihao Sun, Kai Li, Xiaolin Hu

Accurate segmentation of neural structures in Electron Microscopy (EM) images is paramount for neuroscience. However, this task is challenged by intricate morphologies, low signal-to-noise ratios, and scarce annotations, limiting the accuracy and generalization of existing methods. To address these challenges, we seek to leverage the priors learned by visual foundation models on a vast amount of natural images to better tackle this task. Specifically, we propose a novel framework that can effectively transfer knowledge from Segment Anything 2 (SAM2), which is pre-trained on natural images, to the EM domain. We first use SAM2 to extract powerful, general-purpose features. To bridge the domain gap, we introduce a Feature-Guided Attention module that leverages semantic cues from SAM2 to guide a lightweight encoder, the Fine-Grained Encoder (FGE), in focusing on these challenging regions. Finally, a dual-affinity decoder generates both coarse and refined affinity maps. Experimental results demonstrate that our method achieves performance comparable to state-of-the-art (SOTA) approaches with the SAM2 weights frozen. Upon further fine-tuning on EM data, our method significantly outperforms existing SOTA methods. This study validates that transferring representations pre-trained on natural images, when combined with targeted domain-adaptive guidance, can effectively address the specific challenges in neuron segmentation.

Yixiao Li, Xiaoyuan Yang, Weide Liu, Xin Jin, Xu Jia, Yu-Kun Lai, Paul L. Rosin, Hantao Liu, Wei Zhou

As super-resolution (SR) techniques introduce unique distortions that fundamentally differ from those caused by traditional degradation processes (e.g., compression), there is an increasing demand for specialized video quality assessment (VQA) methods tailored to SR-generated content. One critical factor affecting perceived quality is temporal inconsistency, which refers to irregularities between consecutive frames. However, existing VQA approaches rarely quantify this phenomenon or explicitly investigate its relationship with human perception. Moreover, SR videos exhibit amplified inconsistency levels as a result of enhancement processes. In this paper, we propose Temporal Inconsistency Guidance for Super-resolution Video Quality Assessment (TIG-SVQA) that underscores the critical role of temporal inconsistency in guiding the quality assessment of SR videos. We first design a perception-oriented approach to quantify frame-wise temporal inconsistency. Based on this, we introduce the Inconsistency Highlighted Spatial Module, which localizes inconsistent regions at both coarse and fine scales. Inspired by the human visual system, we further develop an Inconsistency Guided Temporal Module that performs progressive temporal feature aggregation: (1) a consistency-aware fusion stage in which a visual memory capacity block adaptively determines the information load of each temporal segment based on inconsistency levels, and (2) an informative filtering stage for emphasizing quality-related features. Extensive experiments on both single-frame and multi-frame SR video scenarios demonstrate that our method significantly outperforms state-of-the-art VQA approaches.

Xuchen Li, Xuzhao Li, Shiyu Hu, Kaiqi Huang, Wentao Zhang

Recent advances in large language models (LLMs) have improved reasoning in text and image domains, yet achieving robust video reasoning remains a significant challenge. Existing video benchmarks mainly assess shallow understanding and reasoning and allow models to exploit global context, failing to rigorously evaluate true causal and stepwise reasoning. We present CausalStep, a benchmark designed for explicit stepwise causal reasoning in videos. CausalStep segments videos into causally linked units and enforces a strict stepwise question-answer (QA) protocol, requiring sequential answers and preventing shortcut solutions. Each question includes carefully constructed distractors based on error type taxonomy to ensure diagnostic value. The benchmark features 100 videos across six categories and 1,852 multiple-choice QA pairs. We introduce seven diagnostic metrics for comprehensive evaluation, enabling precise diagnosis of causal reasoning capabilities. Experiments with leading proprietary and open-source models, as well as human baselines, reveal a significant gap between current models and human-level stepwise reasoning. CausalStep provides a rigorous benchmark to drive progress in robust and interpretable video reasoning.

Shuwei Li, Lei Tan, Robby T. Tan

Day-to-night unpaired image translation is important to downstream tasks but remains challenging due to large appearance shifts and the lack of direct pixel-level supervision. Existing methods often introduce semantic hallucinations, where objects from target classes such as traffic signs and vehicles, as well as man-made light effects, are incorrectly synthesized. These hallucinations significantly degrade downstream performance. We propose a novel framework that detects and suppresses hallucinations of target-class features during unpaired translation. To detect hallucination, we design a dual-head discriminator that additionly performs semantic segmentation to identify hallucinated content in background regions. To suppress these hallucinations, we introduce class-specific prototypes, constructed by aggregating features of annotated target-domain objects, which act as semantic anchors for each class. Built upon a Schrödinger Bridge-based translation model, our framework performs iterative refinement, where detected hallucination features are explicitly pushed away from class prototypes in feature space, thus preserving object semantics across the translation trajectory. Experiments show that our method outperforms existing approaches both qualitatively and quantitatively. On the BDD100K dataset, it improves mAP by 15.5% for day-to-night domain adaptation, with a notable 31.7% gain for classes such as traffic lights that are prone to hallucinations.