Advances in mobile imaging enable smartphones to capture 4K UHD images with rich detail, yet real-world degradations such as moiré patterns, rain streaks, and low-light noise make UHD restoration challenging. Existing methods are often computationally and memory intensive at UHD scale and underutilize degradation-specific cues and internal statistics under resource constraints, limiting faithful detail recovery and global structural consistency. We propose Prior-driven Patch-based Bilateral Compensation (P2BiC), a memory-efficient UHD restoration framework that processes UHD images as patches via an in-network feature-space cropping computation path while maintaining global coherence. Our key insight is that task-specific priors extracted from the observation can exhibit a higher structure-to-artifact ratio than the raw degraded input, whereas the degraded input still provides reliable global tone statistics even under color shifts, which is critical to suppress tone drift and boundary seams in patch-based UHD inference. Accordingly, we introduce prior-guided detail transfer to transfer structure-oriented cues from priors for detail reconstruction and style-aware tone adjustment to align tonal consistency across patches for coherent color. Extensive experiments show that P2BiC achieves strong quality on UHD image demoiréing, deraining, and low-light enhancement, with improved memory scalability.
Robust feature detection and description are core components of 3D vision pipelines, yet they remain challenging under large viewpoint, scale, and appearance changes. Recent methods have improved local feature learning, but many still rely on convolutional operators whose local receptive fields limit their ability to capture long-range geometric cues. We present Robust Deformable Detector (RDD), a robust keypoint detector and descriptor that uses deformable transformers to capture global context and geometric invariance. Deformable attention allows the descriptor branch to focus on a sparse set of informative locations, reducing the effective search space while preserving the ability to model long-range relationships. We also collect an Air-to-Ground dataset for training and evaluation, complementing the standard MegaDepth dataset. RDD outperforms state-of-the-art keypoint detection and description methods in sparse matching and also supports semi-dense matching. To support comprehensive evaluation, we introduce two challenging benchmarks: MegaDepth-View, which emphasizes large viewpoint and scale variations, and Air-to-Ground, which targets cross-altitude matching for 3D reconstruction. Project page: https://xtcpete.github.io/rdd/ .
Transferability of adversarial examples enables black-box attacks without knowledge of victim model parameters and has drawn significant attention. Previous works mostly study gradient variation or image transformations to amplify the distortion on critical parts of inputs. These approaches transfer effectively across similar convolutional neural networks but struggle with heterogeneous architectures, such as vision transformers. Alternatively, model ensemble adversarial attacks have been proposed to fuse outputs from several surrogate models, which can make the generated adversarial example more likely to transfer to other models as it can fool multiple models concurrently. However, existing ensemble attacks generally fuse the outputs of the surrogate models evenly and thus do not explicitly exploit and amplify the intrinsic transfer information of adversarial examples. In this paper, we extend the conference-version AdaEA into AdaEA++. The central methodological advance of AdaEA++ is AGM++, a trajectory-decoupled weighting strategy that estimates input-conditioned surrogate contributions at the initial attack state and reuses them during subsequent optimization, thereby separating model-contribution estimation from the attack trajectory that it induces. Two complementary methodological refinements improve the reliability of this adaptive ensemble: RIE conditions all surrogate models on a shared masked view, while DRF++ replaces pixel-wise disparity estimation with neighborhood-aware gradient consensus. Together, these designs address temporal feedback, spatially unstable evidence, and local gradient misalignment in heterogeneous surrogate ensembles. As a result, AdaEA++ improves the average ASR from 46.90 to 68.16% on NIPS2017 and from 58.68 to 68.80% on the full ImageNet-1K validation set, while maintaining strong transferability on ImageNet-C and ImageNet-A. The proposed AdaEA++ can also boost existing transfer-based attacks, which further demonstrates its efficacy and versatility. The code and data related to our study will be available after publication.
Controllable Human-Object Interaction synthesis aims to generate realistic HOI motions while ensuring all the given control constraints are exactly satisfied. Existing methods primarily generate the entire sequence with all control signals aggressively fused as input conditions in a single stage. However, these methods have significant limitations in satisfying multiple heterogeneous constraints while synthesizing realistic, detailed interactions all at once, which typically results in suboptimal generation performance. In this work, we propose PICI, a novel progressive HOI synthesis framework that generates realistic HOI sequences with different control constraints stage by stage, while keeping the cross-stage invertibility to enhance the generation consistency. Specifically, our PICI includes a sparse HOI generator that produces motion satisfying trajectory constraints and roughly aligning with text semantics, followed by a series of invertible infilling blocks that progressively refine the motion with additional interaction details, conditioned on the knowledge acquired in earlier stages. To better exploit the learned knowledge and facilitate progressive infilling, a Cross-density Knowledge Augmented Memory (CKAMem) is formulated to incrementally aggregate motion features throughout the generation process to facilitate coherent and realistic interactions. Combining these designs, our framework effectively enables the progressive generation of realistic and coherent interactions under multiple control constraints. We conduct extensive experiments on the FullBodyManipulation, 3D-FUTURE and InterAct Datasets, and the results demonstrate that our PICI framework outperforms the current state-of-the-art approaches.
Foundation models like CLIP allow zero-shot transfer on various tasks without additional training data. Yet, the zero-shot performance is less competitive than a fully supervised one. Thus, fine-tuning and ensembling are also commonly adopted to better fit the downstream tasks. However, we argue that such prior work has overlooked the inherent biases in foundation models. Due to the highly imbalanced Web-scale training set, foundation models are inevitably skewed toward frequent semantics, and thus the subsequent fine-tuning or ensembling is still biased. In this study, we systematically examine the biases in foundation models and demonstrate the efficacy of our proposed Generalized Logit Adjustment (GLA) method. Note that bias estimation in foundation models is challenging, as most pre-train data cannot be explicitly accessed like in traditional long-tailed classification tasks. To this end, GLA offers two alternative methods for debiasing: the first is an optimization-based bias estimation built on Bayes optimal criterion, and the second identifies label bias through an eigenvector derived from a matrix of zero-shot predictions. As our work resolves a fundamental flaw in the pre-training, the proposed GLA demonstrates significant improvements across a diverse range of tasks: it achieves 1.5 pp accuracy gains on ImageNet, a large average improvement (1.9−4.4 pp) on 11 few-shot datasets, 2.4 pp gains on long-tailed classification.
