Yuze Wang

Tian Zhang, Yujia Tong, Junhao Dong et al.

The deployment of quantized neural networks on edge devices, combined with privacy regulations like GDPR, creates an urgent need for machine unlearning in quantized models. However, existing methods face critical challenges: they induce forgetting by training models to memorize incorrect labels, conflating forgetting with misremembering, and employ scalar gradient reweighting that cannot resolve directional conflicts between gradients. We propose OEU, a novel Orthogonal Entropy Unlearning framework with two key innovations: 1) Entropy-guided unlearning provides an unbiased forgetting direction by maximizing prediction uncertainty on forgotten data, avoiding confident misprediction toward any specific class, and 2) Gradient orthogonal projection eliminates interference by projecting forgetting gradients onto the orthogonal complement of retain gradients, providing theoretical guarantees for utility preservation under first-order approximation. Extensive experiments demonstrate that OEU outperforms existing methods in both forgetting effectiveness and retain accuracy.

Yuze Wang, Junyi Wang, Chen Wang et al.

This paper introduces a novel continual learning framework for synthesising novel views of multiple scenes, learning multiple 3D scenes incrementally, and updating the network parameters only with the training data of the upcoming new scene. We build on Neural Radiance Fields (NeRF), which uses multi-layer perceptron to model the density and radiance field of a scene as the implicit function. While NeRF and its extensions have shown a powerful capability of rendering photo-realistic novel views in a single 3D scene, managing these growing 3D NeRF assets efficiently is a new scientific problem. Very few works focus on the efficient representation or continuous learning capability of multiple scenes, which is crucial for the practical applications of NeRF. To achieve these goals, our key idea is to represent multiple scenes as the linear combination of a cross-scene weight matrix and a set of scene-specific weight matrices generated from a global parameter generator. Furthermore, we propose an uncertain surface knowledge distillation strategy to transfer the radiance field knowledge of previous scenes to the new model. Representing multiple 3D scenes with such weight matrices significantly reduces memory requirements. At the same time, the uncertain surface distillation strategy greatly overcomes the catastrophic forgetting problem and maintains the photo-realistic rendering quality of previous scenes. Experiments show that the proposed approach achieves state-of-the-art rendering quality of continual learning NeRF on NeRF-Synthetic, LLFF, and TanksAndTemples datasets while preserving extra low storage cost.

Chao Tao, Aoran Hu, Rong Xiao et al.

Data-driven deep learning methods have shown great potential in cropland mapping. However, due to multiple factors such as attributes of cropland (topography, climate, crop type) and imaging conditions (viewing angle, illumination, scale), croplands under different scenes demonstrate a great domain gap. This makes it difficult for models trained in the specific scenes to directly generalize to other scenes. A common way to handle this problem is through the "Pretrain+Fine-tuning" paradigm. Unfortunately, considering the variety of features of cropland that are affected by multiple factors, it is hardly to handle the complex domain gap between pre-trained data and target data using only sparse fine-tuned samples as general constraints. Moreover, as the number of model parameters grows, fine-tuning is no longer an easy and low-cost task. With the emergence of prompt learning via visual foundation models, the "Pretrain+Prompting" paradigm redesigns the optimization target by introducing individual prompts for each single sample. This simplifies the domain adaption from generic to specific scenes during model reasoning processes. Therefore, we introduce the "Pretrain+Prompting" paradigm to interpreting cropland scenes and design the auto-prompting (APT) method based on freely available global land cover product. It can achieve a fine-grained adaptation process from generic scenes to specialized cropland scenes without introducing additional label costs. To our best knowledge, this work pioneers the exploration of the domain adaption problems for cropland mapping under prompt learning perspectives. Our experiments using two sub-meter cropland datasets from southern and northern China demonstrated that the proposed method via visual foundation models outperforms traditional supervised learning and fine-tuning approaches in the field of remote sensing.

Yongtang Bao, Chengjie Tang, Yuze Wang et al.

Reconstructing and segmenting scenes from unconstrained photo collections obtained from the Internet is a novel but challenging task. Unconstrained photo collections are easier to get than well-captured photo collections. These unconstrained images suffer from inconsistent lighting and transient occlusions, which makes segmentation challenging. Previous segmentation methods cannot address transient occlusions or accurately restore the scene's lighting conditions. Therefore, we propose Seg-Wild, an interactive segmentation method based on 3D Gaussian Splatting for unconstrained image collections, suitable for in-the-wild scenes. We integrate multi-dimensional feature embeddings for each 3D Gaussian and calculate the feature similarity between the feature embeddings and the segmentation target to achieve interactive segmentation in the 3D scene. Additionally, we introduce the Spiky 3D Gaussian Cutter (SGC) to smooth abnormal 3D Gaussians. We project the 3D Gaussians onto a 2D plane and calculate the ratio of 3D Gaussians that need to be cut using the SAM mask. We also designed a benchmark to evaluate segmentation quality in in-the-wild scenes. Experimental results demonstrate that compared to previous methods, Seg-Wild achieves better segmentation results and reconstruction quality. Our code will be available at https://github.com/Sugar0725/Seg-Wild.

Yuze Wang, Luo Yang, Junyi Wang et al.

Architecture embodies aesthetic, cultural, and historical values, standing as a tangible testament to human civilization. Researchers have long leveraged virtual reality (VR), mixed reality (MR), and augmented reality (AR) to enable immersive exploration and interpretation of architecture, enhancing accessibility, public understanding, and creative workflows around architecture in education, heritage preservation, and professional design practice. However, existing VR/MR/AR systems are often developed case-by-case, relying on hard-coded annotations and task-specific interactions that do not scale across diverse built environments. In this work, we present ArchGPT, a multimodal architectural visual question answering (VQA) model, together with a scalable data-construction pipeline for curating high-quality, architecture-specific VQA annotations. This pipeline yields Arch-300K, a domain-specialized dataset of approximately 315,000 image-question-answer triplets. Arch-300K is built via a multi-stage process: first, we curate architectural scenes from Wikimedia Commons and filter unconstrained tourist photo collections using a novel coarse-to-fine strategy that integrates 3D reconstruction and semantic segmentation to select occlusion-free, structurally consistent architectural images. To mitigate noise and inconsistency in raw textual metadata, we propose an LLM-guided text verification and knowledge-distillation pipeline to generate reliable, architecture-specific question-answer pairs. Using these curated images and refined metadata, we further synthesize formal analysis annotations-including detailed descriptions and aspect-guided conversations-to provide richer semantic variety while remaining faithful to the data. We perform supervised fine-tuning of an open-source multimodal backbone ,ShareGPT4V-7B, on Arch-300K, yielding ArchGPT.

Yuze Wang, Rong Xiao, Haifeng Li et al.

In remote sensing scene classification, leveraging the transfer methods with well-trained optical models is an efficient way to overcome label scarcity. However, cloud contamination leads to optical information loss and significant impacts on feature distribution, challenging the reliability and stability of transferred target models. Common solutions include cloud removal for optical data or directly using Synthetic aperture radar (SAR) data in the target domain. However, cloud removal requires substantial auxiliary data for support and pre-training, while directly using SAR disregards the unobstructed portions of optical data. This study presents a scene classification transfer method that synergistically combines multi-modality data, which aims to transfer the source domain model trained on cloudfree optical data to the target domain that includes both cloudy optical and SAR data at low cost. Specifically, the framework incorporates two parts: (1) the collaborative transfer strategy, based on knowledge distillation, enables the efficient prior knowledge transfer across heterogeneous data; (2) the information regulation mechanism (IRM) is proposed to address the modality imbalance issue during transfer. It employs auxiliary models to measure the contribution discrepancy of each modality, and automatically balances the information utilization of modalities during the target model learning process at the sample-level. The transfer experiments were conducted on simulated and real cloud datasets, demonstrating the superior performance of the proposed method compared to other solutions in cloud-covered scenarios. We also verified the importance and limitations of IRM, and further discussed and visualized the modality imbalance problem during the model transfer. Codes are available at https://github.com/wangyuze-csu/ESCCS

Weijia Liufu, Xiaoyu Guo, Ruiyi Chen et al.

Vision-Language-Action (VLA) models remain brittle in long-horizon, contact-rich manipulation because success-only imitation provides little supervision for execution drift, while failed rollouts are often discarded. We introduce RePO-VLA, a recovery-driven policy optimization framework that assigns distinct roles to success, recovery, and failure trajectories. RePO-VLA first applies Recovery-Aware Initialization (RAI), slicing recovery segments and resetting history so corrective actions depend on the current adverse state rather than the preceding failure. It then learns a Progress-Aware Semantic Value Function (PAS-VF), aligning spatiotemporal trajectory features with instructions and successful references. The resulting labels salvage useful failure prefixes via reliability decay, while low-value labels mark drift and terminal breakdowns, teaching differences among nominal, failed, and corrective actions. The data engine turns adverse states into planner-generated or human-collected corrective rollouts, teaching recovery to the success manifold. Value-Conditioned Refinement (VCR) trains the policy to prefer high-progress actions. At deployment, a fixed high value ($v=1.0$) biases actions toward the learned success manifold without online failure detectors or heuristic retries. We introduce FRBench, with standardized error injection and recovery-focused evaluation. Across simulated and real-world bimanual tasks, RePO-VLA improves robustness, raising adversarial success from 20% to 75% on average and up to 80% in scaled real-world trials.

Yansong Qu, Shaohui Dai, Xinyang Li et al.

Reconstructing 3D objects from a single image remains challenging, especially under real-world occlusions. While recent diffusion-based view synthesis models can generate consistent novel views from a single RGB image, they typically assume fully visible inputs and fail when parts of the object are occluded, resulting in degraded 3D reconstruction quality. We propose DeOcc-1-to-3, an end-to-end framework for occlusion-aware multi-view generation that synthesizes six structurally consistent novel views directly from a single occluded image, enabling reliable 3D reconstruction without prior inpainting or manual annotations. Our self-supervised training pipeline leverages occluded-unoccluded image pairs and pseudo-ground-truth views to teach the model structure-aware completion and view consistency. Without modifying the original architecture, we fully fine-tune the view synthesis model to jointly learn completion and multi-view generation. Additionally, we introduce the first benchmark for occlusion-aware reconstruction, covering diverse occlusion levels, object categories, and masking patterns, providing a standardized protocol for future evaluation.

Yuze Wang, Yujia Tong, Xuan Liu et al.

Large Language Models (LLMs) inevitably memorize sensitive information during training, posing significant privacy risks. Machine unlearning has emerged as a promising solution to selectively remove such information without full retraining. However, existing methods are designed for dense models and overlook model sparsification, an essential technique for efficient LLM deployment. We find that unlearning effectiveness degrades substantially on sparse models. Through empirical analysis, we reveal that this degradation occurs because existing unlearning methods require updating all parameters, yet sparsification prunes substantial weights to zero, fundamentally limiting the model's forgetting capacity. To address this challenge, we propose Sparsity-Aware Unlearning (SAU), which decouples unlearning from sparsification objectives through gradient masking that redirects updates to surviving weights, combined with importance-aware redistribution to compensate for pruned parameters. Extensive experiments demonstrate that SAU significantly outperforms existing methods on sparse LLMs, achieving effective forgetting while preserving model utility.

Yujia Tong, Yuze Wang, Jingling Yuan et al.

Model quantization enables efficient deployment of deep neural networks on edge devices through low-bit parameter representation, yet raises critical challenges for implementing machine unlearning (MU) under data privacy regulations. Existing MU methods designed for full-precision models fail to address two fundamental limitations in quantized networks: 1) Noise amplification from label mismatch during data processing, and 2) Gradient imbalance between forgotten and retained data during training. These issues are exacerbated by quantized models' constrained parameter space and discrete optimization. We propose Q-MUL, the first dedicated unlearning framework for quantized models. Our method introduces two key innovations: 1) Similar Labels assignment replaces random labels with semantically consistent alternatives to minimize noise injection, and 2) Adaptive Gradient Reweighting dynamically aligns parameter update contributions from forgotten and retained data. Through systematic analysis of quantized model vulnerabilities, we establish theoretical foundations for these mechanisms. Extensive evaluations on benchmark datasets demonstrate Q-MUL's superiority over existing approaches.

Yuze Wang, Yue Qi

Recent advances in leveraging large-scale Internet photo collections for 3D reconstruction have enabled immersive virtual exploration of landmarks and historic sites worldwide. However, little attention has been given to the immersive understanding of architectural styles and structural knowledge, which remains largely confined to browsing static text-image pairs. Therefore, can we draw inspiration from 3D in-the-wild reconstruction techniques and use unconstrained photo collections to create an immersive approach for understanding the 3D structure of architectural components? To this end, we extend language embedded 3D Gaussian splatting (3DGS) and propose a novel framework for open-vocabulary scene understanding from unconstrained photo collections. Specifically, we first render multiple appearance images from the same viewpoint as the unconstrained image with the reconstructed radiance field, then extract multi-appearance CLIP features and two types of language feature uncertainty maps-transient and appearance uncertainty-derived from the multi-appearance features to guide the subsequent optimization process. Next, we propose a transient uncertainty-aware autoencoder, a multi-appearance language field 3DGS representation, and a post-ensemble strategy to effectively compress, learn, and fuse language features from multiple appearances. Finally, to quantitatively evaluate our method, we introduce PT-OVS, a new benchmark dataset for assessing open-vocabulary segmentation performance on unconstrained photo collections. Experimental results show that our method outperforms existing methods, delivering accurate open-vocabulary segmentation and enabling applications such as interactive roaming with open-vocabulary queries, architectural style pattern recognition, and 3D scene editing.

Jinrui Liu, Bingyan Nie, Boyu Li et al.

Improving the reasoning capabilities of embodied agents is crucial for robots to complete complex human instructions in long-view manipulation tasks successfully. Despite the success of large language models and vision language models based on Supervised Fine-Tuning (SFT) in planning tasks, they continue facing challenges in performing long-horizon manipulation tasks in complex real-world environments, owing to their restricted common sense and reasoning capabilities. Considering that aligning general-purpose vision language models to robotic planning tasks via supervised fine-tuning suffers from poor generalization and insufficient physical understanding, we propose RoboGPT-R1, a two-stage fine-tuning framework for embodied planning. In this framework, supervised training acquires foundational knowledge through expert sequences, followed by RL to address the model's shortcomings in visual-spatial understanding and reasoning. To achieve physical understanding and action sequence consistency in multi-step reasoning tasks, we design a rule-based reward function that simultaneously considers long-horizon performance and action constraint in the environment. The reasoning model, trained on Qwen2.5-VL-3B, significantly outperforms the larger-scale model, GPT-4o-mini, by 21.33% and surpasses other work trained on Qwen2.5-VL-7B by 20.33% on the EmbodiedBench benchmark.

Yujia Tong, Tian Zhang, Jingling Yuan et al.

Vision Transformers (ViTs) have revolutionized computer vision tasks with their exceptional performance. However, the introduction of privacy regulations such as GDPR and CCPA has brought new challenges to them. These laws grant users the right to withdraw their data, necessitating not only the deletion of data but also the complete removal of its influence from trained models. Machine unlearning emerges as a critical solution, with exact unlearning being computationally prohibitive and approximate methods offering a more practical approach. This work addresses the particularly challenging scenario of random data forgetting in ViTs, where the model must forget specific samples while retaining others, even within the same class. We first reveal the core characteristics of ViTs through selective masking experiments: when high-attention areas are masked, the model retains its recognition capability but significantly weakens its memorization ability. Based on the above insights, we propose LetheViT, a contrastive unlearning method tailored for ViTs. LetheViT uses masked image inputs to generate positive logits and original image inputs to generate negative logits, guiding the model to forget specific details while retaining the general cl category outlines. Experimental results demonstrate that LetheViT achieves state-of-the-art performance, effectively balancing privacy compliance with model efficacy.

Yuze Wang, Mariana Belgiu, Haiyang Wu et al.

Satellite Image Time Series (SITS) is crucial for agricultural semantic segmentation. However, Cloud contamination introduces time gaps in SITS, disrupting temporal dependencies and causing feature shifts, leading to degraded performance of models trained on complete SITS. Existing methods typically address this by reconstructing the entire SITS before prediction or using data augmentation to simulate missing data. Yet, full reconstruction may introduce noise and redundancy, while the data-augmented model can only handle limited missing patterns, leading to poor generalization. We propose a joint learning framework with feature reconstruction and prediction to address incomplete SITS more effectively. During training, we simulate data-missing scenarios using temporal masks. The two tasks are guided by both ground-truth labels and the teacher model trained on complete SITS. The prediction task constrains the model from selectively reconstructing critical features from masked inputs that align with the teacher's temporal feature representations. It reduces unnecessary reconstruction and limits noise propagation. By integrating reconstructed features into the prediction task, the model avoids learning shortcuts and maintains its ability to handle varied missing patterns and complete SITS. Experiments on SITS from Hunan Province, Western France, and Catalonia show that our method improves mean F1-scores by 6.93% in cropland extraction and 7.09% in crop classification over baselines. It also generalizes well across satellite sensors, including Sentinel-2 and PlanetScope, under varying temporal missing rates and model backbones.

Yuze Wang, Aoran Hu, Ji Qi et al.

Accurately mapping large-scale cropland is crucial for agricultural production management and planning. Currently, the combination of remote sensing data and deep learning techniques has shown outstanding performance in cropland mapping. However, those approaches require massive precise labels, which are labor-intensive. To reduce the label cost, this study presented a weakly supervised framework considering multi-temporal information for large-scale cropland mapping. Specifically, we extract high-quality labels according to their consistency among global land cover (GLC) products to construct the supervised learning signal. On the one hand, to alleviate the overfitting problem caused by the model's over-trust of remaining errors in high-quality labels, we encode the similarity/aggregation of cropland in the visual/spatial domain to construct the unsupervised learning signal, and take it as the regularization term to constrain the supervised part. On the other hand, to sufficiently leverage the plentiful information in the samples without high-quality labels, we also incorporate the unsupervised learning signal in these samples, enriching the diversity of the feature space. After that, to capture the phenological features of croplands, we introduce dense satellite image time series (SITS) to extend the proposed framework in the temporal dimension. We also visualized the high dimensional phenological features to uncover how multi-temporal information benefits cropland extraction, and assessed the method's robustness under conditions of data scarcity. The proposed framework has been experimentally validated for strong adaptability across three study areas (Hunan Province, Southeast France, and Kansas) in large-scale cropland mapping, and the internal mechanism and temporal generalizability are also investigated.

Yuze Wang, Junyi Wang, Yue Qi

Novel View Synthesis (NVS) from unconstrained photo collections is challenging in computer graphics. Recently, 3D Gaussian Splatting (3DGS) has shown promise for photorealistic and real-time NVS of static scenes. Building on 3DGS, we propose an efficient point-based differentiable rendering framework for scene reconstruction from photo collections. Our key innovation is a residual-based spherical harmonic coefficients transfer module that adapts 3DGS to varying lighting conditions and photometric post-processing. This lightweight module can be pre-computed and ensures efficient gradient propagation from rendered images to 3D Gaussian attributes. Additionally, we observe that the appearance encoder and the transient mask predictor, the two most critical parts of NVS from unconstrained photo collections, can be mutually beneficial. We introduce a plug-and-play lightweight spatial attention module to simultaneously predict transient occluders and latent appearance representation for each image. After training and preprocessing, our method aligns with the standard 3DGS format and rendering pipeline, facilitating seamlessly integration into various 3DGS applications. Extensive experiments on diverse datasets show our approach outperforms existing approaches on the rendering quality of novel view and appearance synthesis with high converge and rendering speed.