cost-volume construction

To address the limitations of local stereo matching methods—lacking global consistency—and global methods—suffering from excessive computational cost—this paper proposes an efficient, fine-tuning-free depth estimation model that generalizes across resolutions and disparity ranges. The method introduces a multi-resolution Transformer architecture incorporating sparse attention to significantly reduce memory consumption; a probabilistic focusing loss that jointly optimizes disparity, occlusion, and confidence maps; and eliminates both cost volume filtering and deep refinement networks. Evaluated on Middlebury v3 and ETH3D benchmarks, the model achieves state-of-the-art accuracy—substantially outperforming existing approaches—while maintaining high computational efficiency and superior detail reconstruction capability.

This work addresses the failure of conventional stereo matching and monocular depth estimation under challenging scenarios—including textureless regions, occlusions, specular, and transparent objects—where ground-truth depth supervision is unavailable. We propose the first zero-shot cross-domain stereo matching framework. Methodologically, we design a dual-branch collaborative architecture: a geometric branch enforcing disparity consistency and a vision foundation model (VFM) branch incorporating monocular priors; introduce a geometry-guided cost volume fusion mechanism for cross-modal feature alignment; and construct MonoTrap, an optically plausible synthetic dataset enabling purely synthetic, zero-shot training. Without access to real-world annotations, our method achieves state-of-the-art zero-shot performance across multiple benchmarks, significantly outperforming existing stereo and monocular approaches on challenging cases such as specular and transparent surfaces.

Existing stereo matching models exhibit poor generalization to unseen domains, primarily due to insufficient diversity in training data. Method: We propose a general-purpose stereo matching foundation model for open-world scenarios, introducing a novel hybrid-data scaling paradigm: large-scale synthetic stereo pairs are constructed by fusing real annotated stereo images with monocular-image-based generation, augmented via controllable multi-baseline, multi-view, and multi-scene synthesis. A unified end-to-end architecture is adopted, integrating self-supervised pair generation, cross-domain mixed training, and zero-shot transfer evaluation. Contribution/Results: The model achieves strong zero-shot generalization across five public benchmarks—adapting robustly to novel environments and sensor configurations without fine-tuning. Code is publicly available.

To address the challenge of simultaneously achieving high accuracy, efficiency, and low computational cost in real-time stereo matching, this paper proposes a lightweight 3D cost volume modeling framework. It replaces computationally expensive 4D cost aggregation with a channel-optimized 3D cost volume and, for the first time, systematically exploits the representational capacity along the channel dimension via a multi-strategy channel enhancement mechanism—including channel-wise attention refinement and lightweight feature encoding. The framework further integrates 2D cost aggregation with an efficient spatial regularization network. Despite an extremely low computational load (22 GFLOPs, 17 ms), the model achieves competitive end-point error (EPE) on SceneFlow and ranks first on the KITTI 2015 real-time leaderboard. The core contribution lies in establishing a channel-dimension-driven paradigm for efficient 3D cost volume modeling.

To address the limited accuracy and robustness of multi-view depth estimation under diverse camera configurations (e.g., varying relative poses and lens types), this paper proposes a plug-and-play iterative depth hypothesis pruning framework. Given an arbitrary initial depth map, the method adaptively resamples to generate a set of depth hypotheses and—novelly—introduces contrastive learning into the multi-view depth hypothesis space to learn scale- and configuration-invariant discriminative features. It further integrates multi-view geometric constraints with adaptive metric-space mapping to robustly select the optimal hypothesis. Evaluated on standard benchmarks, the approach significantly improves both depth and surface normal estimation accuracy, consistently outperforming state-of-the-art deep learning-based stereo matching methods.

This study addresses the limitations of insufficient accuracy and structural detail in digital surface models (DSMs) generated from satellite stereo imagery by systematically integrating advanced learning-based stereo matching models—StereoAnywhere, MonSter, and Foundation Stereo—into the satellite photogrammetry pipeline for the first time. By designing image rectification strategies tailored to satellite imaging geometry and incorporating disparity polarity and range constraints, the method enables end-to-end DSM generation. Experiments demonstrate that the proposed approach significantly outperforms conventional algorithms across multiple satellite image datasets, yielding DSMs with richer geometric detail and clearer structural fidelity. The authors release open-source code to support large-scale Earth observation applications. Although challenges remain in complex terrains such as dense vegetation and the mean absolute error (MAE) metric exhibits saturation effects, this work establishes an effective framework for deep learning–driven high-precision 3D reconstruction from satellite imagery.

This study addresses the problem of recovering true correspondences from multiple noisy and independently permuted point clouds in high-dimensional space. While single-view observation exhibits an information-theoretic impossibility threshold—where exact matching becomes infeasible when the signal strength parameter $b < 2$—this work demonstrates for the first time that incorporating multiple views circumvents this fundamental limitation. Leveraging a high-dimensional Gaussian model and tools from random matrix theory, the authors devise a polynomial-time algorithm that, given $K$ views, achieves near-perfect recovery with only $o(n)$ mismatches whenever $b > K/(K-1)$. Notably, with three views, the method enables efficient and accurate matching in the regime $3/2 < b < 2$, which is provably impossible under a single view.

Existing truth-free multi-view 3D consistency evaluation methods are prone to high-score hallucinations in the presence of artifacts, redundant views, or noise, making reliable geometric consistency assessment challenging. This work proposes a decomposable neural metric framework that decouples consistency evaluation into backbone, residual, and aggregation components, and integrates geometric verification signals from COLMAP—such as feature matching, registration, dense support, and reconstruction failure detection—to establish the first controllable robustness benchmark. Experiments demonstrate that the proposed metric significantly outperforms existing approaches on real novel-view synthesis results and human evaluations, achieving up to a fourfold increase in correlation with human judgment over MEt3R and a threefold improvement in robustness. This study also systematically reveals, for the first time, that prominent 3D foundation models—including VGGT, MASt3R, DUSt3R, and Fast3R—are susceptible to geometric hallucinations under irrelevant or degenerate inputs.

This work addresses a critical limitation in real-time stereo matching: conventional coarse-to-fine approaches often suffer irreversible matching failures due to fixed local search ranges, which can miss true correspondences when initial disparity estimates are inaccurate. To overcome this, the authors propose an Uncertainty-Guided Residual Search Mechanism (UGRSM) that dynamically re-centers the local cost volume based on the confidence of propagated disparities within a coarse-to-fine framework, enabling adaptive search. This approach significantly enhances the robustness of local matching, yielding substantial improvements in disparity accuracy across multiple benchmarks—including SceneFlow, KITTI, Middlebury, and ETH3D—while maintaining real-time inference efficiency.

This work addresses the limitations of existing pairwise image matching methods, which often yield fragmented and geometrically inconsistent correspondences in multi-view settings, hindering high-quality 3D reconstruction. To overcome this, we propose MV-RoMa, a geometrically consistent dense multi-view matching model that jointly estimates dense correspondences from a source image to multiple co-visible images, enabling coherent multi-view trajectory reconstruction. Our approach introduces a novel multi-view encoder and a pixel-level attention refinement module, which leverages pairwise matches as geometric priors to enhance consistency while avoiding the high computational cost of full cross-attention. Additionally, we design an SfM-oriented trajectory integration strategy for post-processing. Experiments demonstrate that MV-RoMa significantly outperforms current methods on multiple challenging benchmarks, producing denser, more reliable correspondences and thereby enabling more accurate 3D reconstruction.