SCOPE: Scale-Consistent One-Pass Estimation of 3D Geometry
Abstract Overview
SCOPE is a method for estimating 3D geometry from extended monocular video sequences while trying to preserve both geometric accuracy and long-range temporal consistency. The model predicts affine-invariant 3D point maps for all frames with shared sequence-level scale and shift parameters, so the output remains scale-consistent across the video rather than being solved independently per frame. Its design combines multi-scale geometric supervision, cross-frame geometric constraints in a shared reference frame, hierarchical temporal supervision across multiple time scales, and NTK-adapted rotary positional encoding with sequence-stretching for long-sequence extrapolation. The method is intended to process long videos in a single pass and to support downstream reconstruction pipelines such as MegaSAM.
Novelty
The paper’s main distinction is a one-pass monocular video geometry estimator that explicitly enforces a shared affine ambiguity across an entire sequence instead of predicting framewise geometry independently. It also combines viewpoint-invariant cross-frame geometry, appearance-invariant temporal supervision, and frequency-modulated positional extrapolation to handle sequences much longer than those seen during training.
Results
Across Sintel, ScanNet, Bonn, and KITTI, SCOPE achieves the best average rank among the methods reported in Table 1. The paper reports a 24.2% reduction in relative point-map error and a 34.9% reduction in temporal alignment error on ScanNet versus prior state-of-the-art methods, and Table 2 shows competitive or better accuracy than video-depth baselines while running substantially faster. For long sequences, single-pass inference on 270-frame clips slightly but consistently outperforms sliding-window inference, and qualitative reconstruction results indicate more coherent 4D outputs than MoGe under the same downstream setup.
Key Points
- SCOPE predicts per-frame 3D point maps from monocular video with shared sequence-level scale consistency, aiming to reduce drift over hundreds of frames.
- The method introduces cross-frame geometric losses, multi-scale temporal derivative supervision, and NTK-scaled RoPE with sequence stretching to improve long-range consistency and extrapolation.
- Experiments report stronger temporal alignment and competitive geometry accuracy across several datasets, along with efficient single-pass processing of 300 frames at about 39.1 FPS.