Dense Prediction

Vision Transformers (ViTs) can learn strong image-level representations while their patch representations become less effective for dense prediction during prolonged training. We revisit this dense degradation phenomenon and argue that it is not fully explained by high-norm artifacts alone. Instead, we characterize \emph{semantic diffusion}: an optimization shortcut in which global semantic information spreads through patch tokens beyond what is locally justified. Our analysis shows that dense representation quality is not captured by locality alone: shallow features can remain better aligned with foreground regions yet underperform deeper features, and \texttt{[CLS]} features remain complementary for dense prediction. These observations suggest that the goal should not be to remove global context, but to make token interactions more selective. We therefore study sparse attention as a minimal intervention, replacing softmax attention with entmax-1.5 while preserving global token connectivity. On DINOv1 ViT-S/16 trained for 200 epochs on ImageNet-1K, this change preserves ImageNet linear probing accuracy and substantially improves semantic segmentation performance: VOC mIoU increases from 42.80 to 48.78, ADE20K from 19.85 to 21.97, and Cityscapes from 36.79 to 37.87. These results suggest that selective token mixing is a simple and effective bias for improving dense ViT representations.

Mueller matrix imaging provides rich, physically meaningful contrast for biomedical tissue analysis, but supervised learning is hindered by scarce dense annotations and strong domain shifts across specimens and acquisition settings. We introduce MuellerPT, a physics guided pre-training approach that learns transferable dense representations by predicting Lu-Chipman decomposition maps from per-pixel 4x4 Mueller matrices. To scale pre-training, we collected a new large Multispectral Animal Polarimetric Organ dataset (MAP-Org). The pre-trained encoder is adapted with a segmentation head for grey vs. white matter segmentation in lamb brain. A classification head is used for colorectal cancer vs. non-cancer classification. Both segmentation and classification are evaluated across few-shot learning scenarios. In segmentation, MuellerPT improves label efficiency and cross specimen transfer compared to models without pre-training, achieving an absolute DICE gain of over 20% compared to the baseline trained from scratch when using 5% of the training data. In classification, MuellerPT also enhances label efficiency, improving overall accuracy by 8% compared to the baseline when using 1% of the training data. We demonstrate MuellerPT's robustness to domain shift with a qualitative evaluation of its predicted Lu-Chipman maps on an ex vivo human oesophagus sample. These results suggest that predicting Lu-Chipman decomposition is an effective and practical pretext task for robust biomedical inference from Mueller polarimetry and can pave the way for future work on label efficient Mueller imaging.

Estimating forest aboveground biomass (AGB) from Earth observation combines two structurally incompatible label sources: spaceborne lidar provides canopy structure at millions of locations but no biomass estimate, and ground-based plots provide biomass at thousands of biased locations but no metrics of structure. No single training sample carries labels for all target variables, plot labels are missing not at random (MNAR), and biomass is linked to the structural variables by known but biome-specific allometric laws. We formalise this as multi-task dense regression under heterogeneous disjoint partial supervision with MNAR labels and inter-task physical constraints, and propose StruMPL to address it jointly. A shared encoder feeds per-variable regression, imputation, and propensity heads for spatial MNAR correction, and a learnable physics module that evaluates the inter-task constraint on the model's own predictions at every pixel. The supervised loss uses an Augmented IPW (AIPW) pseudo-outcome with stop-gradients on the propensity and on the imputation baseline; we show analytically and empirically that both are necessary for joint optimisation to recover IPW-weighted stationary points while keeping the loss bounded. On two ecologically distinct biomes, StruMPL outperforms ablation variants and the closest published method on AGB RMSE and bias, with a stratified analysis showing AIPW reduces high-AGB bias by ~54%.

Although large-scale visual foundation models (VFMs) achieve remarkable performance in semantic understanding, they still underperform in instance-aware dense prediction tasks. They exhibit different biases in representation: for instance, promptable segmentation models (e.g., SAM2) focus on fine-grained region boundaries, while self-supervised models (e.g., DINOv3) emphasize object-level structure. This observation highlights the potential of combining complementary features from different VFMs to enhance downstream dense prediction tasks. However, naive multi-VFM fusion seldom leads to reliable gains, and interpretable principles for leveraging their complementary features are still underexplored. In this work, we propose a metric-guided approach that effectively selects and aggregates complementary features from different VFMs based on explicit assessment scores. Specifically, we design a suite of label-free metrics in feature space across two aspects, Structural Coherence and Edge Fidelity, to assess features of VFM encoders. Guided by these scores, we identify complementary edge-strong and structure-strong encoder pairs, and integrate them via a master-auxiliary fusion scheme. This feature fusion requires no complex architectural changes and is trained only in a single stage. Our model shows consistent performance gains across multiple dense prediction tasks compared with the baselines, with better object-level semantics and more accurately localized boundaries. The code is available at {https://github.com/gyc-code/metric-guided-fusion}.

Modern image encoders achieve high generalization by decoupling semantic meaning from resolution, an ability yet to be fully realized in the 3D domain. We investigate the failure of 3D point cloud encoders to achieve similar generalization and find that existing models are highly sensitive to sampling resolution and scale changes, leading to significant performance degradation. This sensitivity is a major bottleneck for real-world deployment in robotics, as it suggests models overfit to specific quantization densities and object scales rather than learning invariant semantic features. To mitigate this dependency, we propose Invaria, a point cloud encoder that achieves scale and density invariance through next-resolution prediction and receptive field calibration. While our objective is not the explicit generation of high-resolution point clouds, we find that this training objective encourages the model to learn robust, structural invariants. The resulting encoder achieves significant performance gains during resolution shifts while maintaining high efficiency through a compact model size and reduced token requirements. Specifically, on ScanNet, Invaria achieves a 56.0% higher mIoU at 3$\times$ lower resolution and a 20% improvement when the objects scale is reduced by a factor of 3. These gains are achieved with a 45% smaller model size and an average reduction of 40% in input tokens.

INTRODUCTION | Fully supervised 3D segmentation of high-resolution ex vivo MRI is limited by the prohibitive cost of volumetric annotation, forcing reliance on sparse 2D slices. Weakly supervised Sparse-to-Dense frameworks bridge this gap, but guidelines remain ambiguous regarding human-centric visual enhancements and transferring optimization strategies across dimensions. We analyze divergent regularization needs for multi-class segmentation of high-resolution ex vivo spinal cord MRI. METHODS | We used 9.4T MRI of multiple sclerosis spinal cords (>104,000 slices) with sparse annotations (428 slices). A 2D Teacher trained on sparse slices generated dense pseudo-labels to train a 3D Student. We systematically evaluated the impact of human-centric preprocessing, spatial augmentation, and soft-label regularization on both architectures. RESULTS | We identified a critical divergence in training dynamics. The 2D Teacher required strong spatial augmentation and soft-labeling to overcome data scarcity, improving White Matter Lesion Dice scores by >11 points. However, propagating these techniques to the 3D Student degraded its performance. Furthermore, human-centric preprocessing (e.g., CLAHE) disrupted global statistical cues, dropping Gray Matter Lesion Dice scores by ~25 points. DISCUSSION | Our study highlights a perception divergence (human-centric contrast enhancement harms machine models) and a regularization conflict across dimensions. 3D architectures trained on dense pseudo-labels exhibit fundamentally different optimization landscapes than 2D counterparts and require distinct, conservative regularization. Code and models: https://github.com/ivadomed/model_seg_sc-gm-lesion_human_ms_exvivo_t2star.

Shadow detection is commonly formulated as a vision-driven dense prediction problem, where models rely primarily on pixel-wise visual supervision to distinguish shadows from non-shadow regions. However, this formulation can become unreliable in visually ambiguous cases, where similar dark regions may correspond either to cast shadows or to intrinsically dark surfaces, making visual evidence alone insufficient for establishing a stable decision rule. In this work, we revisit shadow detection from a vision--language perspective and argue that robust prediction benefits from an explicit semantic reference beyond visual cues alone. We propose SVL, a Shadow Vision--Language framework that uses language as an explicit semantic reference to disambiguate shadows from visually similar dark regions. SVL aligns global image representations with shadow-related text embeddings through scene-level shadow ratio regression, and transfers this semantic guidance to dense prediction via global-to-local coupling and local patch-level constraints. Built on a frozen DINOv3 image encoder, SVL learns only lightweight projection and decoding modules, yielding a parameter-efficient design with less than $1\%$ trainable parameters. Extensive experiments on multiple shadow detection benchmarks, including dedicated hard-case evaluations, suggest strong overall performance and improved robustness under visually ambiguous conditions. Code is available at https://github.com/harrytea/SVL.

Open-vocabulary semantic segmentation requires adapting image-level vision-language models such as CLIP to dense pixel-level prediction, which is challenging due to the mismatch between hierarchical structure and semantic alignment in the embedding space. While recent works leverage hyperbolic geometry to model hierarchical relationships, they align embeddings across hierarchical levels but overlook semantic misalignment among embeddings within the same level. In this work, we propose HyRo, a hyperbolic fine-tuning framework that decouples hierarchical and semantic alignment in the Poincaré ball model. HyRo aligns hierarchical levels by adjusting the hyperbolic radius and refines semantic relationships through angular alignment using an orthogonal transformation that theoretically preserves the hyperbolic radius. Experiments on standard open-vocabulary semantic segmentation benchmarks demonstrate that HyRo achieves state-of-the-art performance over prior methods.

Density estimation and reliable prediction regions for outputs are crucial in supervised and unsupervised learning. While conformal prediction effectively generates coverage-guaranteed regions, it struggles with multi-dimensional outputs due to reliance on one-dimensional nonconformity scores. To address this, we introduce CONTRA: CONformal prediction region via normalizing flow TRAnsformation. CONTRA utilizes the latent spaces of normalizing flows to define nonconformity scores based on distances from the center. This allows for the mapping of high-density regions in latent space to sharp prediction regions in the output space, surpassing traditional hyperrectangular or elliptical conformal regions. Further, for scenarios where other predictive models are favored over flow-based models, we extend CONTRA to enhance any such model with a reliable prediction region by training a simple normalizing flow on the residuals. We demonstrate that both CONTRA and its extension maintain guaranteed coverage probability and outperform existing methods in generating accurate prediction regions across various datasets. We conclude that CONTRA is an effective tool for (conditional) density estimation, addressing the under-explored challenge of delivering multi-dimensional prediction regions.

Recent progress in dense optical flow has been driven by increasingly complex architectures and multi-step refinement for test-time scaling. While these approaches achieve strong benchmark performance, they also require substantial computation during inference. This raises a fundamental question: Is scaling test-time computation the only way to improve dense optical flow accuracy? We argue that it is not. Instead, powerful visual semantic and geometric priors encoded in modern foundation models can reduce, if not overcome, the need for computationally expensive iterative refinement at test-time. In this paper, we present a framework that estimates dense optical flow in a single forward pass, leveraging pretrained foundation representations, while avoiding iterative refinement and additional inference-time computation, thus offering an alternative to test-time scaling. Our method extracts visual semantic features from a frozen DINO-v2 backbone and combines them with geometric cues from a monocular depth foundation model. We fuse these complementary priors into a unified representation and apply a global matching formulation to estimate dense correspondences without recurrent updates or test-time optimization. Despite avoiding iterative refinement, our approach achieves strong cross-dataset generalization across challenging benchmarks. On Sintel Final, we obtain 2.81 EPE without refinement, significantly improving over state-of-the-art (SOTA) SEA-RAFT under comparable training conditions and outperforming RAFT, GMFlow (without refinement), and recent FlowSeek in the same setting. These results suggest that strong foundation priors can substitute for test-time scaling, offering a computationally efficient alternative to refinement-heavy pipelines.

Monocular depth estimation (MDE) with self-supervised training approaches struggles in low-texture areas, where photometric losses may lead to ambiguous depth predictions. To address this, we propose a novel technique that enhances spatial information by applying a distance transform over pre-semantic contours, augmenting discriminative power in low texture regions. Our approach jointly estimates pre-semantic contours, depth and ego-motion. The pre-semantic contours are leveraged to produce new input images, with variance augmented by the distance transform in uniform areas. This approach results in more effective loss functions, enhancing the training process for depth and ego-motion. We demonstrate theoretically that the distance transform is the optimal variance-augmenting technique in this context. Through extensive experiments on KITTI, Cityscapes, Waymo, NYUv2 and ScanNet our model demonstrates robust performance, surpassing competing self-supervised methods in MDE.

We present CoopNet, an approach that improves the cooperation of co-trained networks by dynamically adapting the apportionment of gradient, to ensure equitable learning progress. It is applied to motion-aware self-supervised prediction of depth maps, by introducing a new hybrid loss, based on a distribution model of photo-metric reconstruction errors made by, on the one hand the depth + odometry paired networks, and on the other hand the optical flow network. This model essentially assumes that the pixels from moving objects (that must be discarded for training depth and odometry), correspond to those where the two reconstructions strongly disagree. We justify this model by theoretical considerations and experimental evidences. A comparative evaluation on KITTI and CityScapes datasets shows that CoopNet improves or is comparable to the state-of-the-art in depth, odometry and optical flow predictions.

Monocular depth estimation (MDE) typically produces depth estimations that are defined up to an unknown scale or shift. When only sparse metric anchors are available, recovering accurate metric depth becomes challenging yet necessary for practical applications. We address this problem by formulating metric depth recovery as image-adaptive scale field modeling. Instead of directly correcting the depth, we reformulate the correction as a low-dimensional linear combination of image-adaptive basis maps. These maps are derived from semantic and geometric cues encoded in the MDE estimations and intermediate representations. The weights of basis maps are efficiently determined from sparse metric anchors via a least-squares problem. This formulation yields improved metric depth accuracy, strong robustness under extreme anchor sparsity, and an interpretable decomposition of spatial scale variations. Extensive experiments across multiple datasets and representative MDE models demonstrate the effectiveness and general applicability of our approach.

Accurate Digital Surface Model (DSM) reconstruction from satellite imagery is critical for applications such as disaster response, urban planning, and large-scale geographic mapping. Existing approaches face a fundamental trade-off: optimization-based methods achieve strong accuracy but require hours of per-scene computation, while generalizable geometry foundation models offer near-instant inference but fail to generalize to satellite imagery due to the domain gap introduced by the Rational Polynomial Camera (RPC) model and mismatched depth scale distributions. We present Sat3R, a feed-forward framework that bridges this gap via RPC-aware metric depth fine-tuning of Depth Anything V2 using the Scale-Invariant Logarithmic (SiLog) loss. By constructing physically consistent pseudo depth supervision from RPC geometry, Sat3R adapts a monocular depth foundation model to the satellite domain without per-scene optimization. Experiments on the DFC2019 benchmark demonstrate that Sat3R reduces MAE by 38% over zero-shot feed-forward baselines and achieves competitive accuracy against optimization-based methods, while delivering over 300x speedup. Sat3R demonstrates that feed-forward models, when properly adapted to the satellite domain, can match optimization-based accuracy at a fraction of the computational cost, paving the way for practical large-scale satellite DSM reconstruction.

Cell-level dense prediction is central to computational pathology, but remains challenging due to fine-grained histological structures, strong domain shifts, and costly dense annotations. Existing ViT-based pathology foundation models rely on patch tokenization, which can disrupt spatial continuity and weaken local morphological details needed for cell-level prediction. To address this, we propose Masked-Diffusion Convolutional Foundation Models, termed ConvNeXt Masked-Diffusion (CMD), a self-supervised convolutional generative pretraining framework for dense pathology representation learning. CMD uses a fully convolutional ConvNeXt-UNet backbone, performs masked-diffusion pretraining in pixel space, and incorporates frozen pathology foundation model features through adaptive normalization. Experimental results demonstrate that CMD consistently outperforms existing ViT-based pathology foundation models and even surpasses state-of-the-art end-to-end segmentation methods while fine-tuning only a small number of task-specific parameters across multiple pathology dense prediction tasks. The advantage is particularly pronounced under limited annotation settings, where CMD exhibits stronger robustness and generalization ability. Our findings suggest that purely convolutional architectures can also serve as competitive pathology foundation models for cell-level dense prediction, achieving leading performance within the current ViT-dominated paradigm and providing a scalable, high-performance solution that better preserves histological structural priors for fine-grained pathology understanding.

Multi-task dense prediction solves complementary pixel-level tasks in a unified model, such as semantic segmentation, depth estimation, surface normal estimation, and edge detection. Existing decoder-side interactions use attention, prompts, routing, diffusion, Mamba, or bridge features to exchange task evidence, but most of them organize this evidence implicitly. They usually fuse task features by similarity or affinity, without explicitly modeling that evidence reliability varies across tasks and spatial locations. As a result, unreliable evidence may contaminate the shared representation and intensify negative transfer. We propose $\mathcal{B}^{3}$-Net, a controlled posterior bridge learning framework for multi-task dense prediction. Our method decomposes decoder-side interaction into reliability estimation, posterior bridge construction, and bounded redistribution. The Precision Field Estimator estimates patch-wise evidence precision from task-reference alignment and local variation. The Posterior Bridge Operator builds a precision-weighted posterior bridge through heteroscedastic evidence fusion, yielding a shared state more reliable than uniform or heuristic mixtures. The Contractive Dispatch Operator redistributes the bridge to each task branch through a bounded update, reducing uncontrolled feature injection. Experiments on NYUD-v2, PASCAL-Context, and Cityscapes show that $\mathcal{B}^{3}$-Net achieves competitive or superior trade-offs over representative CNN-, Transformer-, diffusion-, Mamba-, and bridge-feature-based methods. Backbone-matched comparisons and extensive analyses further verify that the gains arise from controlled posterior bridge learning rather than backbone capacity or decoder scale.

Optical satellite image time series are extensively used in many Earth observation applications, including agriculture, climate monitoring, and land surface analysis. However, clouds and swath edges result in irregular sampling along the temporal dimension, limiting continuous monitoring. To address this issue, a growing body of work has focused on temporal densification and reconstruction of satellite image time series, with the objective of filling missing or cloud-contaminated observations within the temporal extent of the available data. While these approaches improve temporal continuity, they are inherently restricted to the reconstruction of the gaps within the observed time periods, and do not address the prediction of future observations. This work proposes a probabilistic deep learning framework for the densification and forecasting of Sentinel-2 time series by generating optical images at arbitrary past or future dates. The approach leverages multimodal satellite data by jointly exploiting Sentinel-2 optical and Sentinel-1 SAR observations. Unlike most existing works, we propose to focus on the uncertainty of the generated images. Experimental results demonstrate effective densification and forecasting, on sparse and temporally misaligned time series.

Vision Transformers (ViTs) achieve state-of-the-art segmentation accuracy but require large training datasets because each layer has unique parameters that must be learned independently. We present RD-ViT, a Recurrent-Depth Vision Transformer that adapts the Recurrent-Depth Transformer (RDT) architecture to dense prediction tasks, supporting both 2D and 3D inputs. RD-ViT replaces the deep stack of unique transformer blocks with a single shared block looped T times, augmented with LTI-stable state injection for guaranteed convergence, Adaptive Computation Time (ACT) for spatial compute allocation, depth-wise LoRA adaptation, and optional Mixture-of-Experts (MoE) feed-forward networks for category-specific specialization. We evaluate on the ACDC cardiac MRI segmentation benchmark in both 2D slice-level and 3D volumetric settings with exclusively real experiments executed in Google Colab. In 2D, RD-ViT outperforms standard ViT at 10% training data (Dice 0.774 vs 0.762) and at full data (0.882 vs 0.872). In 3D, RD-ViT with MoE achieves Dice 0.812 with 3.0M parameters, reaching 99.4% of standard ViT performance (0.817) at 53% of the parameter count. MoE expert utilization analysis reveals that different experts spontaneously specialize for different cardiac structures (RV, MYO, LV) without explicit routing supervision. ACT halting maps show higher compute allocation at cardiac boundaries, and the mean ponder time decreases from 2.6 to 1.4 iterations during training, demonstrating learned computational efficiency. Depth extrapolation enables inference with more loops than training without degradation. All code, notebooks, and results are publicly released.

We present Sapiens2, a model family of high-resolution transformers for human-centric vision focused on generalization, versatility, and high-fidelity outputs. Our model sizes range from 0.4 to 5 billion parameters, with native 1K resolution and hierarchical variants that support 4K. Sapiens2 substantially improves over its predecessor in both pretraining and post-training. First, to learn features that capture low-level details (for dense prediction) and high-level semantics (for zero-shot or few-label settings), we combine masked image reconstruction with self-distilled contrastive objectives. Our evaluations show that this unified pretraining objective is better suited for a wider range of downstream tasks. Second, along the data axis, we pretrain on a curated dataset of 1 billion high-quality human images and improve the quality and quantity of task annotations. Third, architecturally, we incorporate advances from frontier models that enable longer training schedules with improved stability. Our 4K models adopt windowed attention to reason over longer spatial context and are pretrained with 2K output resolution. Sapiens2 sets a new state-of-the-art and improves over the first generation on pose (+4 mAP), body-part segmentation (+24.3 mIoU), normal estimation (45.6% lower angular error) and extends to new tasks such as pointmap and albedo estimation. Code: https://github.com/facebookresearch/sapiens2

Fitting an underlying body model to 3D clothed human assets has been extensively studied, yet most approaches focus on either single-modal inputs such as point clouds or multi-view images alone, often requiring a known metric scale. This constraint is frequently impractical, especially for AI-generated assets where scale distortion is common. We propose OmniFit, a method that can seamlessly handle diverse multi-modal inputs, including full scans, partial depth observations, and image captures, while remaining scale-agnostic for both real and synthetic assets. Our key innovation is a simple yet effective conditional transformer decoder that directly maps surface points to dense body landmarks, which are then used for SMPL-X parameter fitting. In addition, an optional plug-and-play image adapter incorporates visual cues to compensate for missing geometric information. We further introduce a dedicated scale predictor that rescales subjects to canonical body proportions. OmniFit substantially outperforms state-of-the-art methods by 57.1 to 80.9 percent across daily and loose clothing scenarios. To the best of our knowledge, it is the first body fitting method to surpass multi-view optimization baselines and the first to achieve millimeter-level accuracy on the CAPE and 4D-DRESS benchmarks.

Depth completion from sparse LiDAR measurements and corresponding RGB images is a prerequisite for accurate 3D perception in robotic systems. Existing methods achieve high accuracy on standard benchmarks but rely on heavy backbone architectures that preclude real-time deployment on embedded hardware. We present EfficientPENet, a two-branch depth completion network that replaces the conventional ResNet encoder with a modernized ConvNeXt backbone, introduces sparsity-invariant convolutions for the depth stream, and refines predictions through a Convolutional Spatial Propagation Network (CSPN). The RGB branch leverages ImageNet-pretrained ConvNeXt blocks with Layer Normalization, 7x7 depthwise convolutions, and stochastic depth regularization. Features from both branches are merged via late fusion and decoded through a multi-scale deep supervision strategy. We further introduce a position-aware test-time augmentation scheme that corrects coordinate tensors during horizontal flipping, yielding consistent error reduction at inference. On the KITTI depth completion benchmark, EfficientPENet achieves an RMSE of 631.94 mm with 36.24M parameters and a latency of 20.51 ms, operating at 48.76 FPS. This represents a 3.7 times reduction in parameters and a 23 times speedup relative to BP-Net, while maintaining competitive accuracy. These results establish EfficientPENet as a practical solution for real-time depth completion on resource-constrained edge platforms such as the NVIDIA Jetson.

Salient object detection (SOD) requires modeling both long-range contextual dependencies and fine-grained structural details, which remains challenging for convolutional, transformer-based, and Mamba-based state space models. While recent Mamba-based state space approaches enable efficient global reasoning, they often struggle to recover precise object boundaries. In contrast, diffusion models capture strong structural priors through iterative denoising, but their use in discriminative dense prediction is still limited due to computational cost and integration challenges. In this work, we propose DGSSM, a diffusion-guided state space (Mamba) framework that formulates multimodal salient object detection as a progressive denoising process. The framework integrates diffusion structural priors with multi-scale state space encoding, adaptive saliency prompting, and an iterative Mamba diffusion refinement mechanism to improve boundary accuracy. A boundary-aware refinement head and self-distillation strategy further enhance spatial coherence and feature consistency. Extensive experiments on 13 public benchmarks across RGB, RGB-D, and RGB-T settings demonstrate that DGSSM consistently outperforms state-of-the-art methods across multiple evaluation metrics while maintaining a compact model size. These results suggest that diffusion-guided state space modeling is an effective and generalizable paradigm for multimodal dense prediction tasks.

We present a system for automated detection, localization, and scoring of arrow punctures on 40,cm indoor archery target faces, trained on only 48 annotated photographs (5{,}084 punctures). Our pipeline combines three components: a color-based canonical rectification stage that maps perspective-distorted photographs into a standardized coordinate system where pixel distances correspond to known physical measurements; a frozen self-supervised vision transformer (DINOv3 ViT-L/16) paired with AnyUp guided feature upsampling to recover sub-millimeter spatial precision from $32 \times 32$ patch tokens; and lightweight CenterNet-style detection heads for arrow-center heatmap prediction. Only 3.8,M of 308,M total parameters are trainable. Across three cross-validation folds, we achieve a mean F1 score of $0.893 \pm 0.011$ and a mean localization error of $1.41 \pm 0.06$,mm, comparable to or better than prior fully-supervised approaches that require substantially more training data. An ablation study shows that the CenterNet offset regression head, typically essential for sub-pixel refinement, provides negligible detection improvement while degrading localization in our setting. This suggests that guided feature upsampling already resolves the spatial precision lost through patch tokenization. On downstream archery metrics, the system recovers per-image average arrow scores with a median error of 1.8% and group centroid positions to within a median of 4.00,mm. These results demonstrate that frozen foundation models with minimal task-specific adaptation offer a practical paradigm for dense prediction in small-data regimes.

Depth from Defocus (DfD) is the task of estimating a dense metric depth map from a focus stack. Unlike previous works overfitting to a certain dataset, this paper focuses on the challenging and practical setting of zero-shot generalization. We first propose a new real-world DfD benchmark ZEDD, which contains 8.3x more scenes and significantly higher quality images and ground-truth depth maps compared to previous benchmarks. We also design a novel network architecture named FOSSA. FOSSA is a Transformer-based architecture with novel designs tailored to the DfD task. The key contribution is a stack attention layer with a focus distance embedding, allowing efficient information exchange across the focus stack. Finally, we develop a new training data pipeline allowing us to utilize existing large-scale RGBD datasets to generate synthetic focus stacks. Experiment results on ZEDD and other benchmarks show a significant improvement over the baselines, reducing errors by up to 55.7%. The ZEDD benchmark is released at https://zedd.cs.princeton.edu. The code and checkpoints are released at https://github.com/princeton-vl/FOSSA.

LiDAR point clouds are fundamental to various applications, yet the extreme sparsity of high-precision geometric details hinders efficient context modeling, thereby limiting the compression speed and performance of existing methods. To address this challenge, we propose a compact representation for efficient predictive lossless coding. Our framework comprises two lightweight modules. First, the Geometry Re-Densification Module iteratively densifies encoded sparse geometry, extracts features at a dense scale, and then sparsifies the features for predictive coding. This module avoids costly computation on highly sparse details while maintaining a lightweight prediction head. Second, the Cross-scale Feature Propagation Module leverages occupancy cues from multiple resolution levels to guide hierarchical feature propagation, enabling information sharing across scales and reducing redundant feature extraction. Additionally, we introduce an integer-only inference pipeline to enable bit-exact cross-platform consistency, which avoids the entropy-coding collapse observed in existing neural compression methods and further accelerates coding. Experiments demonstrate competitive compression performance at real-time speed. Code will be released upon acceptance. Code is available at https://github.com/pengpeng-yu/FastPCC.

Feed-forward 3D Gaussian Splatting methods enable single-pass reconstruction and real-time rendering. However, they typically adopt rigid pixel-to-Gaussian or voxel-to-Gaussian pipelines that uniformly allocate Gaussians, leading to redundant Gaussians across views. Moreover, they lack an effective mechanism to control the total number of Gaussians while maintaining reconstruction fidelity. To address these limitations, we present F4Splat, which performs Feed-Forward predictive densification for Feed-Forward 3D Gaussian Splatting, introducing a densification-score-guided allocation strategy that adaptively distributes Gaussians according to spatial complexity and multi-view overlap. Our model predicts per-region densification scores to estimate the required Gaussian density and allows explicit control over the final Gaussian budget without retraining. This spatially adaptive allocation reduces redundancy in simple regions and minimizes duplicate Gaussians across overlapping views, producing compact yet high-quality 3D representations. Extensive experiments demonstrate that our model achieves superior novel-view synthesis performance compared to prior uncalibrated feed-forward methods, while using significantly fewer Gaussians.

Panoramic dense-prediction models, spanning semantic segmentation and depth estimation, are typically trained under a strict gravity-aligned assumption. Real-world captures, however, routinely violate it: handheld devices jitter and aerial platforms change attitude, so the camera is rarely upright. Under such 3D reorientation, standard spherical Transformers overfit global latitude cues and collapse. We introduce SO3UFormer, an architecture that learns intrinsic spherical features largely decoupled from the underlying coordinate frame, through three geometric components: (1) removing absolute latitude encoding, which breaks the dependence on the gravity axis; (2) quadrature-consistent spherical attention, which corrects for non-uniform sampling density; and (3) a gauge-aware relative positional bias built from local tangent-plane angles rather than global axes. A logit-space \emph{SO(3)}-consistency regularizer, used only during training, further suppresses residual discretization effects. To benchmark robustness, we introduce Pose35, a variant of Stanford2D3D perturbed by random rotations within $\pm 35^\circ$, and evaluate under a full, arbitrary \emph{SO(3)} stress test. There, the baseline SphereUFormer collapses from 67.53 \emph{mIoU} on Pose35 to 25.26 under the full \emph{SO(3)} test, whereas SO3UFormer reaches 72.03 on Pose35 and retains 70.67 under the same test. Similarly, on a second real-world dataset (Matterport3D) for segmentation and on panoramic depth estimation, SO3UFormer remains essentially rotation-invariant while the gravity-anchored baseline again loses most of its accuracy. Code and models are available at https://github.com/zhuqinfeng1999/SO3UFormer.

The existence of multiple, equally accurate models for a given predictive task leads to predictive multiplicity, where a Rashomon set of models achieve similar accuracy but diverge in their individual predictions. This inconsistency undermines trust in high-stakes applications where we want consistent predictions. We propose three approaches to reduce inconsistency among predictions for the members of the Rashomon set. The first approach is outlier correction. An outlier has a label that none of the good models are capable of predicting correctly. Outliers can cause the Rashomon set to have high variance predictions in a local area, so fixing them can lower variance. Our second approach is local patching. In a local region around a test point, models may disagree with each other because some of them are biased. We can detect and fix such biases using a validation set, which also reduces multiplicity. Our third approach is pairwise reconciliation, where we find pairs of models that disagree on a region around the test point. We modify predictions that disagree, making them less biased. These three approaches can be used together or separately, and they each have distinct advantages. The reconciled predictions can then be distilled into a single interpretable model for real-world deployment. In experiments across multiple datasets, our methods reduce disagreement metrics while maintaining competitive accuracy.

Dense feature matching aims to estimate all correspondences between two images of a 3D scene and has recently been established as the gold standard due to its high accuracy and robustness. However, existing dense matchers still fail or perform poorly for many hard real-world scenarios, and high-precision models are often slow, limiting their applicability. In this paper, we attack these weaknesses on a wide front through a series of systematic improvements that together yield a significantly better model. In particular, we construct a novel matching architecture and loss, which, combined with a curated diverse training distribution, enables our model to solve many complex matching tasks. We further make training faster through a decoupled two-stage matching-then-refinement pipeline, and at the same time, significantly reduce refinement memory usage through a custom CUDA kernel. Finally, we leverage the recent DINOv3 foundation model along with multiple other insights to make the model more robust and unbiased. In our extensive set of experiments, we show that the resulting novel matcher sets a new state-of-the-art, being significantly more accurate than its predecessors. Code is available at https://github.com/Parskatt/romav2

We propose DLTPose, a novel method for 6DoF object pose estimation from RGBD images that combines the accuracy of sparse keypoint methods with the robustness of dense pixel-wise predictions. DLTPose predicts per-pixel radial distances to a set of minimally four keypoints, which are then fed into our novel Direct Linear Transform (DLT) formulation to produce accurate 3D object frame surface estimates, leading to better 6DoF pose estimation. Additionally, we introduce a novel symmetry-aware keypoint ordering approach, designed to handle object symmetries that otherwise cause inconsistencies in keypoint assignments. Previous keypoint-based methods relied on fixed keypoint orderings, which failed to account for the multiple valid configurations exhibited by symmetric objects, which our ordering approach exploits to enhance the model's ability to learn stable keypoint representations. Extensive experiments on the benchmark LINEMOD, Occlusion LINEMOD and YCB-Video datasets show that DLTPose outperforms existing methods, especially for symmetric and occluded objects. The code is available at https://anonymous.4open.science/r/DLTPose_/ .

In this paper, we explore dense voxel streaming for accurate and efficient 3D occupancy prediction. While dense voxel representations offer fine-grained spatial details and streaming paradigm enables efficient temporal processing, naively combining the two introduces key challenges: (i) warping-induced distortions caused by interpolation used for temporal alignment, and (ii) degraded dynamic object representations due to motion misalignment and detail loss in image-to-voxel projection. To address these, we propose StreamOcc, a novel framework that utilizes two aggregation strategies. Specifically, it first refines propagated voxel features to reduce warping artifacts before temporal accumulation, and then selectively injects instance-level query features encoding dynamic-object semantics into the corresponding occupied voxel regions, preserving temporally consistent modeling while strengthening dynamic object representations. Unlocking effective dense voxel streaming, StreamOcc achieves state-of-the-art performance on SurroundOcc-benchmark and Occ3D-nuScenes under real-time constraints, outperforming the prior best methods by +1.3/2.5 and +1.5/2.0 in (overall/dynamic object) mIoU, respectively, while running at 83.3 ms per frame with only 2.8 GB of memory. The project page is available at https://moonseokha.github.io/StreamOcc/.

Depth foundation models (DFMs) offer strong learned priors for 3D perception from single RGB images but lack physical depth cues, leading to ambiguities in metric scale. We introduce metalenses, an emerging class of ultrathin planar optical elements, as a solution to physically encode missing metric depth cues via nanophotonics. In this paper, we bridge the gap between metalens and DFMs to achieve accurate metric monocular depth sensing. In a single monocular shot, our metalens embeds depth-dependent positional shifts into two polarized optical wavefronts. With an input adaptation strategty, we enable direct fine-tuning that aligns a pretrained DFM with the optical signals. To scale the training data, we further develop a comprehensive simulation pipeline that synthesizes metalens responses from RGB-D datasets, incorporating physical factors to minimize the sim-to-real gap. Experiments demonstrate that this approach outperforms both monocular metric depth estimation and depth-from-defocus baselines, showing an effective pathway for accurate monocular metric depth sensing.

Occupancy prediction has garnered increasing attention in recent years for its comprehensive fine-grained environmental representation and strong generalization to open-set objects. Nevertheless, mainstream occupancy prediction methods employ cumbersome voxel features as the scene representation, incurring substantial overheads in both memory and computation. When comparing the occupancy distribution in each spatial dimension, we find that the information entropy of the height dimension is much lower than the other two dimensions that constitute the Bird's Eye View (BEV) plane, which indicates that the height distribution of occupancy is easier to learn and predict. Accordingly, we propose Lightweight Spatial Embedding that can represent complete height information in a more compact way than voxel features, thus significantly enhancing its deployability. First, Single-Channel Occupancy is sampled from the multi-view depth distributions, which is then processed by Spatial-to-Channel mechanism to extract Lightweight Spatial Embeddings of different views by 2D convolution. These embeddings will interact with each other through the Lightweight Cross-View Interaction module to obtain the Unified Embedding, which can directly supplement BEV features with height information. Furthermore, we extract Edge-aware Spatial Embedding and apply Geometric Supervision on Spatial Embeddings, aiming to enhance their capability to represent spatial information. We also propose BEV-CutMix, a feature-level data augmentation strategy, to increase the diversity of the driving scenes. We integrate these innovative components into a pure 2D convolutional model, namely LightOcc. Sufficient experimental results show that LightOcc achieves state-of-the-art performance on multiple benchmarks while demonstrating significant efficiency advantages.

Dense prediction is a fundamental requirement for many medical vision tasks such as medical image restoration, registration, and segmentation. The most popular vision model, Convolutional Neural Networks (CNNs), has reached bottlenecks due to the intrinsic locality of convolution operations. Recently, transformers have been widely adopted for dense prediction for their capability to capture long-range visual dependence. However, due to the high computational complexity and large memory consumption of self-attention operations, transformers are usually used at downsampled feature resolutions. Such usage cannot effectively leverage the tissue-level textural information available only at the full image resolution. This textural information is crucial for medical dense prediction as it can differentiate the subtle human anatomy in medical images. In this study, we hypothesize that Multi-layer Perceptrons (MLPs) are superior alternatives to transformers in medical dense prediction where tissue-level details dominate the performance, as MLPs enable long-range dependence at the full image resolution. To validate our hypothesis, we develop a full-resolution hierarchical MLP framework that uses MLPs beginning from the full image resolution. We evaluate this framework with various MLP blocks on a wide range of medical dense prediction tasks including restoration, registration, and segmentation. Extensive experiments on six public well-benchmarked datasets show that, by simply using MLPs at full resolution, our framework outperforms its CNN and transformer counterparts and achieves state-of-the-art performance on various medical dense prediction tasks.

A key challenge of real-world image super-resolution (SR) is to recover the missing details in low-resolution (LR) images with complex unknown degradations (e.g., downsampling, noise and compression). Most previous works restore such missing details in the image space. To cope with the high diversity of natural images, they either rely on the unstable GANs that are difficult to train and prone to artifacts, or resort to explicit references from high-resolution (HR) images that are usually unavailable. In this work, we propose Feature Matching SR (FeMaSR), which restores realistic HR images in a much more compact feature space. Unlike image-space methods, our FeMaSR restores HR images by matching distorted LR image features to their distortion-free HR counterparts in our pretrained HR priors, and decoding the matched features to obtain realistic HR images. Specifically, our HR priors contain a discrete feature codebook and its associated decoder, which are pretrained on HR images with a Vector Quantized Generative Adversarial Network (VQGAN). Notably, we incorporate a novel semantic regularization in VQGAN to improve the quality of reconstructed images. For the feature matching, we first extract LR features with an LR encoder consisting of several Swin Transformer blocks and then follow a simple nearest neighbour strategy to match them with the pretrained codebook. In particular, we equip the LR encoder with residual shortcut connections to the decoder, which is critical to the optimization of feature matching loss and also helps to complement the possible feature matching errors. Experimental results show that our approach produces more realistic HR images than previous methods. Codes are released at https://github.com/chaofengc/FeMaSR.

Dense prediction tasks in surgical computer vision, such as segmentation and surgical zone prediction, can provide valuable guidance for laparoscopic and robotic surgery. However, these models often suffer from distribution shifts, as training datasets rarely cover the variability encountered during deployment, leading to poor generalization. We propose DenseTRF, a self-supervised representation adaptation framework based on texture-centric attention. Our method leverages slot attention to learn texture-aware representations that capture invariant visual structures. By adapting these representations to the target distribution without supervision, DenseTRF significantly improves robustness to domain shifts. The framework is implemented through conditioning dense prediction on slot attention and model merging strategies. Experiments across multiple surgical procedures demonstrate improved cross-distribution generalization in comparison to state-of-the-art segmentation models and test-distribution adaptation methods for dense prediction tasks.

Depth completion aims to predict a dense depth map from a color image with sparse depth measurements. Although deep learning methods have achieved state-of-the-art (SOTA), effectively handling the sparse and irregular nature of input depth data in deep networks remains a significant challenge, often limiting performance, especially under high sparsity. To overcome this limitation, we introduce the Gaussian Belief Propagation Network (GBPN), a novel hybrid framework synergistically integrating deep learning with probabilistic graphical models for end-to-end depth completion. Specifically, a scene-specific Markov Random Field (MRF) is dynamically constructed by the Graphical Model Construction Network (GMCN), and then inferred via Gaussian Belief Propagation (GBP) to yield the dense depth distribution. Crucially, the GMCN learns to construct not only the data-dependent potentials of MRF but also its structure by predicting adaptive non-local edges, enabling the capture of complex, long-range spatial dependencies. Furthermore, we enhance GBP with a serial & parallel message passing scheme, designed for effective information propagation, particularly from sparse measurements. Extensive experiments demonstrate that GBPN achieves SOTA performance on the NYUv2 and KITTI benchmarks. Evaluations across varying sparsity levels, sparsity patterns, and datasets highlight GBPN's superior performance, notable robustness, and generalizable capability.