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

Conference: ECCV2026
Paper: ECCV paper page
Code: https://u7079256.github.io/LinStereo/
Area: 3D Vision
Keywords: stereo matching / iterative refinement / linear attention / monocular depth prior / underwater depth estimation

TL;DR¶

LinStereo replaces the ConvGRU update operator of iterative stereo matching — which only ever reasons over a local window — with a Position-Aware Linear Attention (PALA) operator that aggregates global context over the whole feature map at \(O(N)\) cost per step, and pairs it with per-scale Hierarchical Semantic Cost Volumes (HSCV) and a monocular Depth Prior Initialization (DPI); trained on SceneFlow alone it reaches state-of-the-art-level accuracy on standard benchmarks and the best overall accuracy on TartanAir-UW / SQUID without any underwater training data.

Background & Motivation¶

The mainstream turn in stereo matching happened after RAFT-Stereo. Explicit cost-volume methods (PSMNet, GA-Net and their lineage) build a 3D/4D volume and regularize it with 3D convolutions; they are accurate, but cost grows with the disparity range and the volume is built at a single resolution, so multi-level semantics never enter. Iterative methods took a different route: fix one correlation volume and let a recurrent operator query it repeatedly, pushing disparity from zero to its final value over a controllable number of steps. That flexibility made iteration the dominant paradigm, and two lines of work pushed its ceiling higher — swapping conventional encoders for vision foundation model (VFM) backbones (FoundationStereo, DEFOM-Stereo, MGStereo, Stereo Anywhere) for zero-shot generalization, and fusing monocular depth priors to constrain ill-posed regions.

Yet a growing mismatch has appeared: backbones are getting bigger and their features more global, while the update operator has stayed the same. Three information pathways go to waste. First, updates reason over a local spatial neighborhood, so a piece of evidence needs tens of iterations to diffuse across the image — exactly the wrong behavior for textureless and occluded regions that need context from distant reliable pixels. Second, the VFM supplies a multi-scale hierarchy, but correlation volumes are usually built at one fixed resolution (say 1/4), collapsing fine geometric detail and coarse semantic structure into a single level of matching evidence. Third, the geometric priors the backbone already produces are essentially unused at initialization, so refinement still starts from zero disparity and spends its first iterations merely recovering the gross layout. Prior work patches individual pieces — MGStereo adds monocular priors, IGEV swaps in a geometry encoding volume — but nobody has touched the narrow-bandwidth update interface itself.

This paper's judgment is that, since the backbone already provides semantically rich and scale-consistent features, what needs redesigning is not the input but the update operator. Core idea: replace local recurrence with position-aware linear attention so that every refinement step performs global aggregation at linear cost — attention operates globally over the spatial axis while per-scale correlation volumes supply matching evidence and monocular depth supplies a warm start — thereby propagating reliable disparity into degraded regions within far fewer iterations.

Method¶

Overall Architecture¶

LinStereo takes a rectified stereo pair and outputs a dense disparity map. The whole framework is built on Depth Anything V3 (DA3) with its weights fully frozen: DA3 consists of a DINOv2 ViT-B/14 encoder and a Dual-DPT head, where the Dual-DPT head first reassembles the multi-scale intermediate ViT representations through shared modules and then routes them to separate depth and ray fusion branches. LinStereo draws two things from a single forward pass — multi-scale features taken after the shared reassembly layers but before the fusion branches (downsampling factors \(s\in\{4,8,16\}\), projected to a uniform 128 channels) as matching evidence, and the monocular depth map from the depth fusion branch as initialization. The decoder is therefore "one backbone, two pathways, three modifications": HSCV decides what to match, DPI decides where to start, and PALA decides how to refine. The three are not independent drop-in modules but the input and initial-value designs built around the single PALA operator — the paper stresses this explicitly.

Concretely, the rectified pair goes through the frozen DA3 to produce left/right features at scales \(s=4,8,16\) plus an affine-invariant monocular depth map. HSCV computes a 1D correlation volume along the epipolar line at each scale and reorganizes each volume into a 4-level pyramid along the disparity axis — independent of spatial resolution, so the search range grows while spatial detail is preserved. DPI fits a scale-shift transform from sparse SIFT matches and converts monocular depth into a metric disparity used as the round-0 estimate. Refinement then starts: each of the three scales has its own PALA updater, and hidden states are updated coarse-to-fine, \(h_{16}\rightarrow h_8\rightarrow h_4\), with adjacent scales exchanging hidden states; only the finest-scale state \(h_4\) is decoded into a disparity residual. Every refinement step is supervised, with later iterations weighted more heavily. Structurally this is "encode once + three pathways + one loop", not a cascaded multi-stage network.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Rectified stereo pair"] --> B["Frozen DA3 backbone<br/>DINOv2 ViT-B + Dual-DPT"]
    B --> C["Multi-scale VFM features<br/>s = 4 / 8 / 16"]
    B --> D["Affine-invariant monocular depth"]
    C --> E["Hierarchical Semantic Cost Volumes<br/>per-scale volume + disparity pyramid"]
    D --> F["Depth Prior Initialization<br/>sparse matching + scale-shift alignment"]
    E --> G["Position-Aware Linear Attention<br/>per-scale global linear-attention update"]
    F --> G
    G -->|"T iterations, coarse-to-fine h16→h8→h4"| H["Disparity update head → dense disparity"]

Key Designs¶

1. Hierarchical Semantic Cost Volumes (HSCV): giving every scale semantically aligned matching evidence

The pain point is direct: prior iterative methods (RAFT-Stereo, IGEV) build a correlation volume at one fixed resolution, flattening the VFM's hierarchy into single-level correlation and artificially narrowing the semantic bandwidth of matching; the alternative of fusing multiple scales (e.g. pooling them into one feature) consumes the coarse levels only once. HSCV improves on both axes. In utilization, it builds a separate correlation volume for each \(s\in\{4,8,16\}\) and keeps the full set simultaneously accessible, so PALA can re-query any scale at any iteration instead of walking a one-shot coarse-to-fine cascade. In representation, each per-scale volume is internally downsampled along the disparity axis into a 4-level pyramid, giving "coarse disparity range + fine spatial resolution" within a scale — coarse and fine matching share the same spatial resolution rather than being obtained by downsampling the image.

Mechanically, each scale's left/right features are first projected to a uniform channel count \(C_f\) by instance-normalized residual blocks plus 1×1 convolutions, then correlated along the epipolar line:

\[C_s(i,j,d)=\frac{1}{\sqrt{C_f}}\left\langle \hat{\mathbf{F}}_{L,s}(i,j),\ \hat{\mathbf{F}}_{R,s}(i,j-d)\right\rangle\]

where \(i,j\) index spatial positions and \(d\) the disparity displacement. The result is a two-level matching hierarchy: across semantic scales (the VFM's hierarchy) and within each scale (the disparity pyramid). Its value is that every refinement level receives semantically appropriate matching signals at its own operating resolution, without relying on hidden-state propagation between scales — which is both slow and unreliable. The ablation confirms this: replacing per-scale volumes with a single pooled multi-scale feature degrades KITTI 2015 EPE from 1.01 to 1.07, so what matters is per-scale alignment, not multi-scale fusion.

2. Depth Prior Initialization (DPI): turning affine-invariant monocular depth into a metric disparity warm start

Using monocular priors for initialization is not new (MonSter, Stereo Anywhere), but the common choices are either to import the absolute scale of a metric monocular model (Depth Pro, Metric3D v2) or to leave initialization untouched and still start from zero disparity. The former's absolute scale is a statistical prior that drifts badly out of distribution (underwater, for instance); the latter wastes early iterations on recovering the gross layout. LinStereo's choice is to keep the affine-invariant model's robust cross-domain relative structure and let the calibrated stereo pair supply the metric scale, since the rectified pair already provides geometric correspondences from which that scale is directly recoverable.

The mechanism is to first map the monocular prediction from the depth domain to the disparity domain through \(d=\alpha/\hat{D}_{\text{mono}}+\beta\) — disparity is inversely proportional to depth, so \(\alpha\) absorbs the unknown metric scale and \(\beta\) an additive offset — and then fit \(\alpha,\beta\) by least squares over SIFT keypoint pairs \(\{(p_k,d^{\text{sp}}_k)\}\) matched along epipolar lines on the rectified pair:

\[(\alpha^*,\beta^*)=\arg\min_{\alpha,\beta}\sum_{k=1}^{K}\left\|d^{\text{sp}}_k-\left(\frac{\alpha}{\hat{\mathbf{D}}_{\text{mono}}(p_k)}+\beta\right)\right\|^2,\qquad \mathbf{d}^{(0)}=\frac{\alpha^*}{\hat{\mathbf{D}}_{\text{mono}}}+\beta^*\]

SIFT rather than a learned matcher is deliberate: only sparse correspondences are needed, handcrafted features suffice, and no extra network is introduced. When inliers fall below \(K_{\min}=20\) the method falls back to zero-disparity initialization (triggered on 0% of TartanAir-UW frames and 3.7% of SQUID frames, costing only +0.08 px EPE). The authors are candid that \(\mathbf{d}^{(0)}\) is only a coarse warm start: VFM depths do not strictly follow a global affine model and handcrafted matches carry localization noise, so fine-grained disparity still depends on the subsequent PALA iterations.

3. Position-Aware Linear Attention (PALA): asymmetric 2D RoPE restores spatial awareness to linear attention

This is the paper's core operator, replacing the ConvGRU. The motivation: stereo refinement fundamentally needs to propagate reliable depth from textured regions into neighboring textureless ones, yet a local recurrent operator sees only a small neighborhood per step, so crossing the image takes tens of iterations; softmax attention could reason globally but costs \(O(N^2)\), which is prohibitive on high-resolution feature maps.

Linear attention reduces the cost to \(O(N)\) by exploiting the associativity of matrix multiplication — instead of an \(N\times N\) attention matrix it precomputes the key-value product \(K^{\top}V\) once and shares it across all query positions. But that very sharing creates a new problem: all pairwise spatial relationships collapse into one global summary, which makes the attention position-blind. For disparity refinement this is especially damaging — nearby pixels generally share similar depth, and disparity discontinuities align with object boundaries — exactly the structures that a position-blind operator cannot distinguish.

PALA restores spatial awareness by incorporating Rotary Position Embeddings (RoPE) without giving up linear complexity, extended to both the height and width axes (2D RoPE). Given kernel-mapped queries and keys \(\hat{Q}=\phi(W_Qx)\) and \(\hat{K}=\phi(W_Kx)\) with a non-negative kernel \(\phi\), the position-aware variants \(\tilde{Q}=\mathrm{RoPE}(\hat{Q})\), \(\tilde{K}=\mathrm{RoPE}(\hat{K})\) give

\[\mathbf{O}_{\text{attn}}=\frac{\tilde{\mathbf{Q}}\left(\tilde{\mathbf{K}}^{\top}\mathbf{V}\right)}{\hat{\mathbf{Q}}\left(\overline{\hat{\mathbf{K}}}^{\top}\mathbf{1}\right)},\qquad \overline{\hat{\mathbf{K}}}=\frac{1}{N}\sum_j \hat{\mathbf{K}}_j\]

(⚠️ The normalizing term in the denominator is garbled in the extracted text; this reconstruction follows "rotated \(Q,K\) in the numerator, unrotated mean \(\hat{K}\) in the denominator" — refer to the original paper for the exact form.) The key point is that RoPE is applied asymmetrically: only the numerator uses the position-augmented \(\tilde{Q},\tilde{K}\) to encode relative spatial relationships, while the denominator keeps the original \(\hat{Q},\hat{K}\) for stable normalization. Applying RoPE to both would make the normalizing factor position-dependent, distorting attention magnitudes and destabilizing training — the least conspicuous but most consequential engineering judgment in the paper, and the ablation bears it out: symmetric RoPE gives KITTI 2015 EPE 1.05 and TartanAir-UW AbsRel 0.042, both worse than the asymmetric 1.01 / 0.04. Complexity remains \(O(N\cdot C_h^2)\), linear in image resolution.

Two further positional signals are added. Because stereo disparity fields are locally smooth (neighboring pixels on the same surface share similar depth), local spatial encoding on the value branch reinforces neighborhood coherence; absolute position encoding flanking the block anchors features to their spatial coordinates so the network can distinguish systematic depth variation across image regions. Finally, since PALA sits inside an iterative loop, a gated update controls information flow across iterations:

\[\mathbf{z}_g=\sigma\!\left(\mathrm{Conv}_{3\times3}\left([\mathbf{h},\mathbf{O}_{\text{attn}}]\right)\right),\qquad \mathbf{h}^{(t+1)}=(1-\mathbf{z}_g)\odot \mathbf{h}^{(t)}+\mathbf{z}_g\odot \tanh(\cdot)\]

(⚠️ The content wrapped in tanh on the update branch is garbled in the extracted text; refer to the original paper.) The gate is needed because stereo evidence is inherently non-uniform: textured regions yield strong, unambiguous matches while occluded and textureless areas produce noisy signals. With \(\mathbf{z}_g\), each location can aggressively absorb new evidence where matching is confident and preserve its accumulated estimate where it is not, letting reliable disparity gradually propagate into uncertain regions across iterations. LinStereo uses three PALA updaters, one per scale, each containing a single PALA block; hidden states are updated coarse-to-fine with adjacent scales exchanging states, and only the finest scale produces the disparity residual.

A Worked Example¶

Take a reef patch in SQUID whose color has been attenuated into near-texturelessness (\(T=8\), scales 4/8/16). DPI first maps DA3's monocular depth into disparity space through the scale-shift formula: if 20 or more SIFT inliers can be matched around that region, a pair \((\alpha^*,\beta^*)\) is fitted and yields an initial disparity of roughly the right magnitude; if inliers are too few (which happens for about 3.7% of SQUID frames), the method falls back to zero disparity and relies on later iterations. In iteration 1, the coarsest scale 16 performs PALA first: the correlation volume is looked up from HSCV at that scale (and across its disparity pyramid), and global linear attention aggregates the disparity estimates of "well-textured, reliably matched" regions from across the entire image — something a local ConvGRU cannot do, because those regions lie far outside its receptive field. At the gate, the reef pixel's matching confidence is low, so \(\mathbf{z}_g\) is small and it prefers to keep the previous estimate rather than be dragged away by noise; textured pixels, with large \(\mathbf{z}_g\), absorb the new evidence quickly. Scale 16 then passes its hidden state to scale 8, and scale 8 to scale 4, each level receiving semantically aligned correlation at its own resolution. The loop runs eight times; as iterations proceed, disparity from reliable regions seeps into textureless areas layer by layer while the finest-scale residuals polish that coarse estimate into a dense disparity field with crisp boundaries. This also explains why occluded regions improve the most: Middlebury (H) occlusion EPE drops from the previous best of 2.11 to 1.33 (−37%), precisely the region that benefits directly from global propagation.

Loss & Training¶

The finest-scale hidden state \(h_4^{(t+1)}\) is decoded into a disparity residual \(\Delta d^{(t)}\) by stacked depth-wise convolutional blocks. The objective supervises every refinement step with exponentially increasing weights on later iterations:

\[\mathcal{L}=\sum_{t=1}^{T}\gamma^{\,T-t}\Big(\left\|\mathbf{d}^{(t)}-\mathbf{d}_{\text{gt}}\right\|_1+\lambda_s\mathcal{L}_{\text{smooth}}(\mathbf{d}^{(t)})+\lambda_g\mathcal{L}_{\text{grad}}(\mathbf{d}^{(t)})\Big)\]

Here \(\mathcal{L}_{\text{smooth}}\) is an edge-aware smoothness term that penalizes disparity gradients only where image intensity variation is low, encouraging piece-wise smooth predictions while preserving discontinuities at object boundaries; \(\mathcal{L}_{\text{grad}}\) compares predicted and ground-truth disparity gradients by \(\ell_1\) difference at multiple spatial scales to guide accurate depth edges. Both auxiliary terms carry small weights.

In practice the DA3 backbone stays frozen and only the decoder is trained: AdamW, 200K steps on SceneFlow, batch size 8, \(320\times640\) crops, a single H100, one-cycle schedule peaking at \(2\times10^{-4}\), \(T=8\), \(\gamma=0.9\), \(\lambda_s=0.001\), \(\lambda_g=0.01\). Training data is SceneFlow only (about 35K stereo pairs), with no real-world, underwater or domain-specific data at any stage; for real-time use the iteration budget can drop to \(T=2\). The paper additionally contributes SeaStereo-Dataset, a physically rendered underwater stereo corpus (about 40K pairs: ShapeNetCore foreground objects over coral/fish/shipwreck backgrounds under varying Jerlov water types), and a controlled laboratory water-tank benchmark collected in-house at close range (<2 m); details live in the supplementary material.

Key Experimental Results¶

Main Results¶

On standard benchmarks every method uses its officially released weights and is evaluated on the training splits with public ground truth (Middlebury at half resolution, H). LinStereo deliberately stays a ViT-B, SceneFlow-only model to isolate the decoder's contribution, so FoundationStereo (ViT-L with substantial extra data) leading on these benchmarks is expected. In the table, "best same-data baseline" means the best method without extra training data.

Dataset	Metric	Ours (ViT-B, SceneFlow only)	Best same-data baseline	Note
KITTI 2015	EPE / bad-3	1.01 / 4.54	MonSter 0.89 / 3.58	ViT-B on the same table as ViT-L; loses only on EPE
KITTI 2012	EPE / bad-3	0.76 / 3.49	MonSter 0.75 / 3.58	Best bad-3; EPE on par with MonSter
Middlebury (H) all	EPE / bad-2	0.83 / 6.01	DEFOM-Stereo 0.84 / 5.91	Lowest all-region EPE
Middlebury (H) occluded	EPE / bad-2	1.33 / 8.69	DEFOM-Stereo 2.11 / 19.34	37% lower occlusion EPE — the largest single gain
ETH3D	EPE / bad-1	0.24 / 2.41	MonSter 0.23 / 1.41	The one clear shortfall (bad-1 2.41 vs BridgeDepth 1.33)
Booster (Q)	EPE / bad-2	2.14 / 9.93	Stereo Anywhere 2.21 / 9.91	Transparent/specular; 0.02 bad-2 gap is statistical noise
TartanAir-UW	AbsRel / RMSE	0.04 / 2.08	DEFOM-Stereo 0.05 / 3.13	RMSE 31% lower than FoundationStereo (3.01)
SQUID (real-world underwater)	AbsRel / RMSE	0.04 / 0.90	RAFT-Stereo 0.07 / 1.25	Even beats IGEV++ (0.06 / 1.11) with extra data
Laboratory tank (<2 m)	AbsRel / RMSE	0.04 / 0.07	IGEV++ 0.05 / 0.12	Best across the board at close-range real underwater

On the two underwater tables (Tab. 2), LinStereo — trained only on SceneFlow — achieves the best results on both TartanAir-UW and SQUID, surpassing Selective-Stereo, IGEV-Stereo, IGEV++ and FoundationStereo, all of which use real-world or domain-specific data. The largest gains track each benchmark's dominant degradation: long-range backscatter on TartanAir-UW (RMSE lead) and color attenuation on SQUID (AbsRel lead). ⚠️ The abstract claims 28% lower AbsRel on TartanAir-UW and 26% on SQUID, but the tabulated values (0.05→0.04 and 0.06→0.04) imply roughly 20% and 33%, and these do not fully line up with the text's 31% RMSE / 26% AbsRel either; refer to the original paper.

Ablation Study¶

(a) Progressive PALA design (Tab. 4, KITTI 2015 / TartanAir-UW, \(T=8\)):

Config	KITTI EPE / bad-3	TartanAir-UW AbsRel / RMSE	Note
(A) global linear attention	1.18 / 5.32	0.052 / 2.55	position-agnostic baseline
(B) (A) + global spatial encoding	1.12 / 5.05	0.047 / 2.38	largest single-step gain
(C) (B) + local spatial encoding	1.06 / 4.76	0.043 / 2.22	reinforces neighborhood coherence
(D) (C) + adaptive gating (Full PALA)	1.01 / 4.54	0.04 / 2.08	full configuration
(D) but with symmetric RoPE	1.05 / 4.76	0.042 / 2.18	why asymmetric RoPE matters

(b) Component combinations (Tab. 5a, same evaluation protocol):

HSCV	DPI	PALA	KITTI EPE / bad-3	TartanAir-UW AbsRel / RMSE
			1.42 / 6.58	0.068 / 3.21
✓			1.33 / 6.14	0.061 / 2.91
	✓		1.32 / 6.02	0.059 / 2.85
		✓	1.26 / 5.78	0.055 / 2.64
✓	✓		1.21 / 5.51	0.052 / 2.52
✓		✓	1.10 / 4.95	0.045 / 2.28
	✓	✓	1.13 / 5.12	0.047 / 2.35
✓	✓	✓	1.01 / 4.54	0.04 / 2.08
pooled	✓	✓	1.07 / 4.82	0.043 / 2.21

(c) Efficiency (Tab. 6 / Tab. 7, 480×640, single RTX 4500 24 GB):

Method	Params (M)	GFLOPs	Time (ms)	FPS
LightStereo-S	3.44	45.4	9.9	101.0
CoEx	2.73	70.6	11.0	91.0
RAFT-Stereo†	9.87	296.4	18.0	55.7
RT-IGEV	4.17	354.2	24.7	40.5
Ours (\(T=2\))	127.0	770.4	80	12.5

Per-iteration update-operator latency (Tab. 7): PALA 3.50 ± 0.05 ms versus 3.63 ± 0.06 ms for RAFT-Stereo's ConvGRU and 3.43 ± 0.03 ms for IGEV's ConvGRU — one round of global linear attention costs the same as a local ConvGRU.

Key Findings¶

PALA is the single largest contributor, consistent with it being the main architectural change: adding it alone drops KITTI EPE from 1.42 to 1.26, and among pairwise combinations HSCV + PALA is second only to the full model (1.10), showing global attention benefits most from per-scale aligned matching signals.
The three components are super-additive: the sum of individual gains falls clearly short of the 1.42 → 1.01 achieved together, because PALA's global reasoning needs HSCV's per-scale correlations to have semantically aligned evidence to work with, while DPI's warm start lets refinement converge in fewer iterations.
Restoring position awareness is what rescues linear attention: adding global spatial encoding (A→B) is the largest single step in the chain, and within positional encoding, using RoPE asymmetrically beats the symmetric variant (1.05 → 1.01).
Iterations saturate around 8: \(T=12\) and \(T=16\) give essentially no gain over \(T=8\) (EPE 1.03 / 1.02 vs 1.01), while \(T=2\) degrades to 1.24 — global aggregation converges faster, which is also the justification for the real-time \(T=2\) setting.
Per-scale alignment beats multi-scale pooling: swapping HSCV for a single pooled multi-scale feature degrades EPE from 1.01 to 1.07, confirming that what matters is each level receiving correlation aligned with its own resolution, not merely using multiple scales.
Gains concentrate in degraded regions: occlusion (Middlebury H occlusion EPE −37%), transparency/specularity (Booster Q on par with the best) and underwater (AbsRel 0.04) — exactly the cases where photometric cues are unreliable, matching the "propagate from reliable to degraded regions" explanation. Conversely, on ETH3D bad-1, where photometric cues are plentiful and edge detail dominates, it trails (2.41 vs 1.33).

Highlights & Insights¶

Asymmetric RoPE: the most reusable lesson for applying linear attention to dense prediction — put rotary position encoding in the numerator to recover spatial structure, keep the denominator unrotated to preserve normalization stability. Any "linear attention + needs spatial structure" setting (optical flow, disparity, deformable registration) can adopt this directly.
Efficiency along a different axis: prior efficient stereo work cuts the architecture (pruning the disparity search space, replacing 3D aggregation with 2D, dropping explicit volumes); this paper leaves the architecture intact and instead makes each iteration reach further so fewer are needed. The absolute numbers in Tab. 6 (770 GFLOPs, 80 ms) are not cheap — the backbone holds 94% of parameters — so this "efficiency" means reaching equal accuracy in fewer iterations, which is not directly comparable to the lightweight line.
The right way to use a monocular prior: rather than chasing the absolute scale of a metric monocular model (which drifts out of distribution), keep the affine-invariant representation's robust cross-domain structure and let the calibrated stereo pair solve for scale from a few dozen SIFT inliers with a two-parameter least-squares fit. Tiny cost, geometrically plausible start, and a graceful fallback when inliers are scarce.
A deliberate "fix the backbone" choice: pinning to ViT-B and SceneFlow-only puts the comparison on equal footing and makes "how much does the decoder buy?" an answerable question; scaling to ViT-L or more data should continue to help (the authors state the design is backbone-agnostic).

Limitations & Future Work¶

Inference cost remains high: the frozen DA3 accounts for about 94% of the 127 M parameters (roughly 120 M) and most of the runtime; even at \(T=2\) it reaches only 12.5 FPS at 480×640 on an RTX 4500, which makes full-budget deployment on resource-constrained platforms difficult. The authors propose distilling the backbone into a lightweight architecture. ⚠️ Latency is reported only for the \(T=2\) setting that matches the 80 ms in Tab. 6; no full \(T=8\) latency/FPS is given, so the real-time behavior of the full-accuracy setting cannot be read off directly.
DPI depends on handcrafted sparse matching: SIFT provides only a coarse alignment and its localization noise enters the initialization; in textureless or heavily degraded scenes inliers can be insufficient (3.7% of SQUID frames fall back to zero disparity). The authors suggest a lightweight learned sparse matcher as future work.
Underwater evaluation metrics and protocol: TartanAir-UW and SQUID provide depth rather than disparity ground truth, so depth-domain metrics (AbsRel, RMSE) are used and are not directly comparable to disparity EPE; the laboratory tank section scores only AprilTag-valid ground-truth pixels, leaving larger background errors of some baselines outside the quantitative evaluation — the authors state this explicitly, but it means Tab. 3's advantage should be read together with the qualitative results.
A small consistency blemish: the percentage reductions quoted in the abstract and text do not match the tabulated values (see the note above). The direction of the main conclusions is unaffected, but the tables are the authoritative record.

vs RAFT-Stereo / IGEV-style iterative methods: they fix a correlation volume and query it repeatedly with a local ConvGRU, so information spreads by accumulating iterations; IGEV additionally introduces a geometry encoding volume and often pairs it with extra training data. This paper keeps the "iterate + query a correlation volume" paradigm and only swaps the update operator for position-aware linear attention, so each update covers the whole image and converges in fewer iterations (with per-iteration latency on par with a ConvGRU). The price is a more involved update operator (RoPE, gating, dual encoding branches).
vs FoundationStereo / DEFOM-Stereo / Stereo Anywhere / MGStereo: these work on the input side — stronger VFM backbones, monocular priors, zero-shot robustness — but the update interface stays local, creating the "backbone capacity vs. update reach" mismatch. This paper targets exactly that link and deliberately compares at equal backbone/data scale to show the gain comes from the decoder; on standard benchmarks it still yields to a ViT-L + extra-data FoundationStereo, which is a difference in comparison conditions rather than a methodological weakness.
vs efficient stereo methods (LightStereo / CoEx / CGI-Stereo / Fast-ACVNet): they prune the search space and swap in lightweight aggregation to reach tens of GFLOPs and 80–100 FPS, but fall behind clearly in degraded scenes such as underwater (at \(T=2\) this paper's AbsRel 0.05 while the strongest efficient baseline stays at ≥0.07). The two routes are not mutually exclusive — combining HSCV/DPI with a lightweight backbone is the natural next step.

Rating¶

Novelty: ⭐⭐⭐⭐ Brings linear attention and its position-awareness redesign into the update operator of iterative stereo matching, with a genuinely useful observation on asymmetric RoPE; however, each of the three components (VFM multi-scale correlation, monocular warm start, linear attention) has precedent, so the contribution is mainly the combination and the operator swap.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Five standard benchmarks plus two underwater benchmarks plus an in-house laboratory tank, with zero-shot results, an iteration-budget sweep, component ablations and an operator-latency comparison; loses a star for omitting \(T=8\) latency and for the abstract/table percentage mismatch.
Writing Quality: ⭐⭐⭐⭐ The motivation-to-mechanism chain is clear and the three components are precisely positioned; formula typesetting and a few numerical statements are inconsistent between sections.
Value: ⭐⭐⭐⭐ "Fewer iterations for the same accuracy" is a reusable direction in iterative depth estimation, and both asymmetric RoPE and the scale-shift warm start transfer directly to optical flow and multi-view tasks.