UniScale: Unified Scale-Aware 3D Reconstruction for Multi-View Understanding via Prior Injection for Robotic Perception
| Authors | Mohammad Mahdavian et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2602.23224 |
| Download | |
| Categories | cs.CV, cs.RO |
Abstract
We present UniScale, a unified, scale-aware multi-view 3D reconstruction framework for robotic applications that flexibly integrates geometric priors through a modular, semantically informed design. In vision-based robotic navigation, the accurate extraction of environmental structure from raw image sequences is critical for downstream tasks. UniScale addresses this challenge with a single feed-forward network that jointly estimates camera intrinsics and extrinsics, scale-invariant depth and point maps, and the metric scale of a scene from multi-view images, while optionally incorporating auxiliary geometric priors when available. By combining global contextual reasoning with camera-aware feature representations, UniScale is able to recover the metric-scale of the scene. In robotic settings where camera intrinsics are known, they can be easily incorporated to improve performance, with additional gains obtained when camera poses are also available. This co-design enables robust, metric-aware 3D reconstruction within a single unified model. Importantly, UniScale does not require training from scratch, and leverages world priors exhibited in pre-existing models without geometric encoding strategies, making it particularly suitable for resource-constrained robotic teams. We evaluate UniScale on multiple benchmarks, demonstrating strong generalization and consistent performance across diverse environments. We will release our implementation upon acceptance.
Engineering Breakdown
Plain English
UniScale is a neural network system that takes multiple camera views of a scene and reconstructs its 3D structure while simultaneously figuring out camera parameters and the actual physical scale of the environment—a critical capability for robots that need to navigate real spaces. The system works as a single feed-forward network that jointly estimates camera intrinsics and extrinsics (where cameras are and how they're configured), depth maps that don't depend on scale, point cloud representations, and the metric scale of the entire scene. The key innovation is combining global contextual reasoning with camera-aware feature representations to recover absolute metric scale, which prior methods struggled with because depth estimation is inherently scale-ambiguous from 2D images alone. This addresses a concrete robotics problem: accurate 3D scene understanding requires knowing both geometry and scale, but most vision pipelines output geometry that's only correct up to an unknown scaling factor.
Core Technical Contribution
UniScale's core contribution is a modular, unified architecture that solves the metric scale ambiguity problem in multi-view 3D reconstruction by injecting geometric priors through a semantically-informed design. Rather than treating camera parameter estimation, depth prediction, and scale recovery as separate problems, the paper proposes a single jointly-trained network that reasons about all three simultaneously, with camera-aware feature representations that explicitly condition depth and scale reasoning on the camera geometry. The architectural innovation is the flexible integration mechanism for auxiliary geometric priors—the system can optionally incorporate additional constraints or information when available without requiring retraining, making it modular and practical. This differs from prior work which either ignores metric scale (producing only relative geometry), requires separate explicit scale estimation steps, or needs domain-specific manual calibration.
How It Works
The input to UniScale is a sequence of RGB images from multiple viewpoints of the same scene. The network processes these images through a shared encoder that builds camera-aware feature representations—encodings that explicitly encode information about each camera's intrinsic and extrinsic parameters. In parallel, the system computes global contextual features that reason about the entire scene structure to inform scale prediction. The network then produces three main outputs: (1) camera intrinsics and extrinsics for each view, establishing the geometric relationship between cameras, (2) scale-invariant depth and point maps that capture scene geometry up to an unknown multiplicative factor, and (3) a predicted metric scale that converts the scale-invariant geometry into real-world metric coordinates. The key technical mechanism is that global contextual reasoning constrains the scale—by understanding scene semantics and structure globally, the network can infer what the absolute scale must be, solving what would otherwise be an underconstrained problem. The system also includes optional connection points where geometric priors (like known object dimensions or scene constraints) can be injected to improve both robustness and accuracy without retraining.
Production Impact
For robotics teams building autonomous navigation or manipulation systems, UniScale eliminates a major pain point: the need for manual metric scale calibration or the use of fiducial markers to establish absolute scale. Currently, most visual SLAM and 3D reconstruction pipelines output geometry that's accurate in shape but unknown in scale, requiring downstream heuristics (like assuming floor dimensions or using depth sensors for scale anchoring). With UniScale, a robot could use only monocular or stereo cameras and get immediately metrically correct 3D maps suitable for path planning, obstacle avoidance, and manipulation. The production pipeline simplifies: feed multi-view images directly to the model, get back camera poses and metric point clouds, and pass directly to planning algorithms. Trade-offs include the need for multi-view data (not fully optimized for single-image or real-time streaming scenarios yet), training data requirements to learn robust priors, and inference latency for processing multiple image sequences, though the feed-forward architecture (no iterative optimization) suggests reasonable computational cost compared to traditional bundle adjustment methods.
Limitations and When Not to Use This
UniScale assumes sufficient visual overlap and texture across multiple views to recover scale—in textureless or highly repetitive environments (glass walls, blank corridors), the scale ambiguity may not resolve properly without strong priors, making it less suitable for certain industrial or clean-room robotics settings. The paper's abstract doesn't specify performance on small objects or extreme scale variations, suggesting scale recovery may work better for scenes with intermediate-scale structure; extreme close-ups or wide outdoor scenes might pose challenges. The system requires training on datasets with ground-truth metric scale annotations, which are less abundant than relative geometry datasets, potentially limiting generalization to novel domains without fine-tuning. Additionally, the modularity for incorporating geometric priors is mentioned but not detailed in the abstract—it's unclear how robust the system is to missing or incorrect priors, or how much performance depends on having access to domain-specific constraints during deployment.
Research Context
UniScale builds on the established line of work in structure-from-motion and multi-view geometry (estimating 3D structure from image sequences) but tackles the scale ambiguity problem that has plagued monocular depth estimation for years. The work is positioned within modern learning-based 3D vision, where neural networks learn geometric priors from data rather than relying purely on traditional photogrammetry. It advances beyond prior methods that either output only relative geometry (scale-ambiguous) or require explicit two-stage approaches (geometry then scale estimation), moving toward unified end-to-end prediction. The 2026 publication date suggests this represents the current frontier in robotic perception, where the field is moving toward tightly integrated systems that solve multiple related problems simultaneously rather than as separate pipelines.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::