LinStereo: Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching

"The whole is more than the sum of its parts." --- Aristotle

LinStereo: Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching

Yiran Wang^1†, Oliver Turner¹, Viorela Ila^1‡

¹Australian Centre for Robotics, The University of Sydney

^†Corresponding author.   ^‡Project lead.

🌐 Website | 📄 arXiv | 📑 PDF | 🌊 SeaStereo Dataset (soon) | 📌 BibTeX

TL;DR — LinStereo redesigns the iterative stereo update loop around Position-Aware Linear Attention (PALA), a global O(N) operator on a frozen Depth Anything V3 backbone, so reliable disparity propagates across the whole image in a single step. It is competitive on standard benchmarks and generalizes zero-shot to new domains, including underwater. Accepted to ECCV 2026.

Note

🚧 Code, weights, and the SeaStereo dataset are not released yet — they are coming to this repo soon. Star ⭐ / watch 👀 to be notified.

---

📰 News

• 2026.06 📄 Preprint is live on arXiv — arXiv:2606.25437.
• 2026.06 🎉🎉 LinStereo is accepted to ECCV 2026!

Coming soon (TODO):

• Pretrained weights — release the SceneFlow-trained LinStereo checkpoint (frozen Depth Anything V3, ViT-B).
• SeaStereo dataset — release the ~40K-pair physically-rendered underwater corpus with dense disparity ground truth.

---

✏️ Citation

If you find our work helpful, please consider starring ⭐ this repo and citing:

@inproceedings{wang2026linstereo,
  title={LinStereo: Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching},
  author={Wang, Yiran and Turner, Oliver and Ila, Viorela},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026},
  eprint={2606.25437},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

---

🏃 Intro

LinStereo is a general deep stereo-matching method. Iterative stereo updaters rely on local operators (ConvGRUs), so reliable matches diffuse only a few pixels per step and struggle wherever local photometric cues collapse — textureless, non-Lambertian, or light-scattering (underwater) scenes. LinStereo replaces the local updater with a global, linear-complexity one and reads geometry priors from a frozen monocular foundation model.

Key components:

• 🧭 Position-Aware Linear Attention (PALA) — a global attention update operator with O(N) complexity in the number of pixels. Reliable disparity propagates across the whole image in a single step, at roughly the per-iteration cost of a local ConvGRU.
• 🧱 Hierarchical Semantic Cost Volume (HSCV) — multi-scale cost construction guided by foundation-model semantics.
• 🎯 Depth Prior Initialization (DPI) — initializes the iterative loop from the backbone's monocular depth prior for faster, more stable convergence.
• ❄️ Frozen Depth Anything V3 backbone — no backbone fine-tuning; trained on SceneFlow only.
• 🌍 Zero-shot generalization — strong on standard benchmarks and unseen domains (real-world & synthetic underwater) with no domain-specific data.

---

⚡ Quick Start

🚧 Code release is coming soon. The snippet below shows the intended API and will be updated when the code lands in this repo.

# (planned API — not yet available)
import torch
from linstereo import LinStereo

model = LinStereo.from_pretrained("u7079256/LinStereo").eval().cuda()

left  = load_image("left.png")    # (B, 3, H, W)
right = load_image("right.png")
disparity = model(left, right, iters=8)   # (B, 1, H, W)

---

⚖️ Weights

🚧 Pretrained weights will be released here.

Model	Backbone	Training data	Link
LinStereo	Depth Anything V3 · ViT-B/14 (frozen)	SceneFlow only	🚧 Coming soon

---

📦 SeaStereo Dataset

We release SeaStereo, a physically-rendered underwater stereo corpus with dense disparity ground truth.

Stereo pairs	Jerlov water types	Configurations	Ground truth
~40K	7	1000+	dense disparity

Rendering pipeline: ShapeNetCore foreground objects composited over real marine backgrounds (coral, fish, shipwrecks) and rendered in Blender under varying Jerlov water types.

🚧 Download coming soon.

---

📐 PALA vs. ConvGRU

PALA is a global update operator, yet its per-iteration latency is on par with the local ConvGRUs it replaces (480×640, single NVIDIA RTX 4500):

Update operator	Latency (ms) ↓
PALA (Ours)	3.50 ± 0.05
RAFT-Stereo ConvGRU	3.63 ± 0.06
IGEV ConvGRU	3.43 ± 0.03

LinStereo is a large, accuracy-first model (frozen Depth Anything V3 backbone) rather than a lightweight real-time network — see the website for the full efficiency table.

---

🧪 Zero-Shot Results

All methods below are trained on SceneFlow (IGEV++ uses extra training data). Numbers mirror the paper; the full tables with all baselines and best/second-best highlighting are on the website.

Standard benchmarks — EPE

Method	KITTI'15	KITTI'12	Midd(H) Occ	ETH3D	Booster(Q)
RAFT-Stereo	1.13	0.90	3.31	0.36	4.18
MGStereo	1.13	0.87	2.89	0.25	2.26
Stereo Anywhere	1.07	0.83	2.67	0.24	2.21
IGEV++ (+extra)	1.27	1.20	6.83	0.35	5.00
LinStereo (Ours)	1.01	0.76	1.33	0.24	2.14

LinStereo's 1.33 occluded-EPE on Middlebury is 37% below the previous best (DEFOM, 2.11).

Underwater — AbsRel ↓ / RMSE ↓

Method	TartanAir-UW AbsRel	TartanAir-UW RMSE	SQUID AbsRel	SQUID RMSE
RAFT-Stereo	0.08	4.36	0.07	1.25
MGStereo	0.08	3.69	0.09	1.99
Stereo Anywhere	0.06	3.24	0.07	1.46
IGEV++ (+extra)	0.09	4.37	0.06	1.11
LinStereo (Ours)	0.04	2.08	0.04	0.90

Best on every metric on both benchmarks — 28% lower AbsRel and 31% lower RMSE on TartanAir-UW, 26% lower AbsRel on SQUID.

Real-world lab tank — AbsRel ↓ / RMSE ↓ / A1 ↑

Method	AbsRel	RMSE	A1
RAFT-Stereo	0.06	0.15	0.94
MGStereo	0.08	0.18	0.93
Stereo Anywhere	0.09	0.20	0.92
IGEV++ (+extra)	0.05	0.12	0.96
LinStereo (Ours)	0.04	0.07	0.98

Close-range (< 2 m) water tank with AprilTag + CAD-model ground truth to sub-millimetre, including ~3 mm taut ropes as a fine-structure stress test.

---

😘 Acknowledgement

• We thank Jay Zhang and Dr. Gideon Billings for help collecting the real-world underwater data.
• LinStereo builds on a frozen Depth Anything V3 backbone, and our evaluation follows recent stereo work — RAFT-Stereo, IGEV / IGEV++, Stereo Anywhere, DEFOM-Stereo, MGStereo.

---

📜 License

The code, pretrained weights, and SeaStereo dataset will be released for non-commercial research use (CC BY-NC-SA 4.0). Final license terms will accompany the code release.