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.
Figures & tables
Fig. 1 : UniScale Overview. Upon receiving a set of images with optional camera intrinsic and extrinsic information, UniScale generates depth and point maps, metric-scale and auxilliary camera information, all of which may be used in 3D reconstruction for downstream robotic tasks.
Fig. 2 : (a) Overview of the UniScale architecture and (b) Architecture of the Scale Head. The model combines global contextual information from class tokens, camera intrinsics and extrinsics encoded in camera tokens, and image features from aggregated patch tokens to predict the scene-level scale value.
KITTI
ScanNet
Approach
K
Poses
rel ↓
τ↑
rel ↓
τ↑
a) Multi-View Metric
MAST3R [ 9 ]
✗
✗
61.4
0.4
12.80
19.4
MUSt3R [ 8 ]
✗
✗
19.76
7.3
7.66
35.7
MapAnything [ 15 ]
✗
✗
5.67
42.7
32.26
7.1
UniScale
✗
✗
5.19
49.6
5.68
44.5
TABLE I : Comparison with SOTA methods on the Robust-MVD benchmark [ 52 ] using KITTI [ 53 ] and ScanNet [ 50 ] . We evaluate (a) multi-view metric prediction (lower is better) and (b) multi-view with alignment (lower is better for rel, higher is better for τ ), under different combinations of known intrinsics ( K ) and poses.
Fig. 3 : Comparison between UniScale and other SOTA methods on modified dense- N -view benchmark. UniScale demonstrates better or comparable dense multi-view reconstruction for number of input views varying from 2 to 50.
TABLE II : Ablation of the Metric Scale Head on Robust-MVD. We evaluate different scale-head designs for UniScale on KITTI [ 53 ] and ScanNet [ 50 ] to analyze the contribution of each component.
KITTI
ScanNet
Approach/Setting
K
Poses
rel ↓
τ↑
rel ↓
τ↑
Multi-View Metric
a) Images-only
No Prior Injection
✗
✗
5.19
48.6
5.70
44.8
No Prior Injection into Scale Head
✗
✗
5.12
49.6
6.61
38.9
UniScale
✗
✗
5.19
49.6
5.68
44.5
TABLE III : Prior Injection Ablation on Robust-MVD [ 52 ] . We evaluate UniScale with and without prior injection and scale head as well as variants that exclude priors from the scale head, on KITTI [ 53 ] and ScanNet [ 50 ]
Fig. 6 : Comparison between 6D and Quaternion Encoding Strategies. 6D extrinsic encodings demonstrate better results on multi-view reconstruction for number of input views varying from 2 to 50 and under different input configurations.