[CVPR 2025] DAGE: Breaking the Resolution-Length Trade-off in 3D Video Geometry

TL;DR

Reconstructing accurate 3D geometry from uncalibrated videos has historically forced a compromise: you either get global consistency (at the cost of blurry, low-res results) or high-resolution detail (at the cost of temporal jitter and "broken" 3D structures). DAGE (Dual-stream Architecture for Geometry Estimation) solves this by using a low-res "context" stream and a high-res "detail" stream fused by a clever adapter. It supports 2K resolution and clips up to 1000 frames, running up to 28x faster than previous state-of-the-art models like Pi3.

The "Resolution Wall" in 3D Reconstruction

Current feed-forward 3D foundation models (like VGGT and Pi3) rely on Global Attention. Mathematically, this is elegant—it allows every pixel in every frame to "talk" to every other pixel. However, the quadratic complexity of attention ($O(N^2\cdot H{W}^2)$) creates a hardware wall.

If you want to keep the geometry consistent across 100 frames, you have to shrink the images to ~500px. The result? Small objects vanish, and edges become a blurry mess. If you use a single-image model (like DepthPro) at 2K, it looks great per-frame, but the camera poses are missing, and the 3D points don't align across time, creating "layering artifacts."

Methodology: The Power of Disentanglement

DAGE’s core insight is that global coherence doesn't require high resolution. You can figure out where the camera is and how the scene is laid out using a tiny, downsampled version of the video.

The Dual-Stream Pipeline

1. Low-Resolution (LR) Stream: Processes frames at $\le 252px$. It uses alternating frame-global attention to build a view-consistent map and regress camera poses ($SE(3)$).
2. High-Resolution (HR) Stream: Processes the original image at native resolution (up to 2K) per-frame using a frozen ViT backbone (MoGe2). This preserves "pixel-perfect" edges.
3. The Adapter: This is the glue. It injects the global context from the LR stream into the HR stream.

Figure 1: DAGE Architecture showing the decoupled LR (Global) and HR (Local) pathways.

The "Snapping" RoPE Trick

Fusing a 252px feature map with a 2K feature map is non-trivial. Standard Positional Encodings (RoPE) fail when you move to 2K because the model hasn't "seen" such high coordinate values. DAGE uses Interpolated RoPE and a "Snapping" mechanism: each HR token (at 2K) is mapped to its nearest grid cell in the LR map (at 252px), ensuring the cross-attention remains stable and resolution-agnostic.

Experiments & Results

DAGE was tested across 18 diverse datasets, including UrbanSyn and KITTI.

Performance & Efficiency

DAGE doesn't just match previous models; it enables hardware-limited tasks to run at production speeds. At 2K resolution:

• Pi3: 0.2 FPS (and often crashes due to Out-of-Memory).
• DAGE: 5.6 FPS (a 28x speedup).

Visual Fidelity

Traditional models often suffer from "Correspondence Collapse" at high resolutions—attention maps become diffuse and multi-modal. DAGE avoids this by keeping the expensive matching in the low-res domain.

Figure 2: Qualitative 3D reconstruction comparison. DAGE preserves sub-centimeter details that VGGT and Pi3 miss.

Critical Insight: Why This Matters

The "brute force" scaling of transformers is hitting a ceiling in spatial tasks. DAGE represents a shift toward Heuristic Hybridization: using the best properties of single-image foundation models (zero-shot detail) and video models (temporal consistency).

By decoupling the "camera stream" from the "geometry stream," we can finally process 4K drone footage or long cinematic shots without needing a cluster of H100s.

Conclusion & Limitations

DAGE is a significant step toward practical, real-time 3D reconstruction from any video. However, it still struggles with Extremely Rapid Motion where the LR stream may fail to find matches, and it currently assumes a Static Scene (it doesn't recover the moving parts of dynamic objects).

For the 3D Vision community, DAGE's "Adapter" approach provides a blueprint for how to upgrade existing low-res video models to the high-res era.