---
title: 'StereoAdapter: Adaptation in Stereo Vision'
url: https://www.emergentmind.com/topics/stereoadapter
type: topic
---

# StereoAdapter: Adaptation in Stereo Vision

Searching arXiv for papers related to "StereoAdapter" and closely named variants to ground the article in current literature.
StereoAdapter is a name that has been used for multiple, technically distinct research artifacts in stereo vision and related imaging systems. In the literature, it denotes at least four major lines of work: a planar-mirror smartphone catadioptric stereo rig for 3D reconstruction [2109.11872]; a meta-learning and online-adaptation framework for deep stereo, later extended to federated adaptation and geometry-aware domain adaptation [1904.02957], [2405.14873], [2103.14333]; a parameter-efficient stereo image super-resolution method based on adapter modules inserted into a frozen single-image transformer [2407.03598]; and an underwater stereo depth estimation framework that combines LoRA-adapted monocular foundation encoders with recurrent stereo refinement, followed by a second generation using a selective state-space updater [2509.16415], [2602.16915]. The shared label therefore refers not to a single canonical architecture, but to a family of adapter-oriented or adaptation-oriented approaches that address stereo acquisition, stereo matching, stereo super-resolution, and underwater stereo depth estimation.

## 1. Terminological scope and disambiguation

The term “StereoAdapter” first appears, in the supplied corpus, in “Learning to Adapt for Stereo,” where it refers to a framework that enables deep stereo methods to continuously adapt to new target domains in an unsupervised manner [1904.02957]. In that work, the emphasis is on online parameter adaptation of a stereo network through a learned unsupervised objective weighting mechanism.

A different use appears in “Catadioptric Stereo on a Smartphone,” where StereoAdapter is a compact, low-cost catadioptric rig that converts the front- and back-facing cameras of an off-the-shelf smartphone into a calibrated stereo pair [2109.11872]. Here the term denotes a physical optical-mechanical device rather than a learning algorithm.

The name is later used again in distinct algorithmic contexts. In “ASTeISR,” “stereo adapters” are small cross-view fusion modules and “spatial adapters” are local refinement modules inserted into a frozen single-image super-resolution transformer for stereo image super-resolution [2407.03598]. In “StereoAdapter: Adapting Stereo Depth Estimation to Underwater Scenes,” the term refers to a self-supervised underwater stereo depth estimation system with LoRA-adapted monocular priors and a recurrent stereo refinement module [2509.16415]. “StereoAdapter-2” then replaces the ConvGRU updater with a ConvSS2D operator based on selective state-space models and augments the training data with UW-StereoDepth-80K [2602.16915].

This multiplicity of meanings suggests that “StereoAdapter” functions as a recurrent research label for stereo-specific adaptation mechanisms. A plausible implication is that any technical discussion must distinguish between physical stereo conversion, online domain adaptation, parameter-efficient transfer, and underwater stereo estimation, because their assumptions, objectives, and evaluation protocols differ substantially.

## 2. Smartphone catadioptric StereoAdapter

In “Catadioptric Stereo on a Smartphone,” StereoAdapter is implemented as a 3D-printed mount (“mask”) and two planar mirrors, each \(3\,\mathrm{cm}\times 3\,\mathrm{cm}\), which steer light from the scene into the front and rear smartphone cameras [2109.11872]. The two mirrors are held at an identical tilt angle \(\beta\) with respect to the phone’s body, one in front of the rear camera and one in front of the selfie camera. By reflection, each real camera produces an image of a virtual camera whose optical center lies behind the mirror. The effective baseline is \(b \simeq 2\cdot b_m\), where \(b_m\) is the distance from the phone’s camera lens to the mirror plane.

The optical design is expressed in terms of mirror side length \(l_m\), mirror tilt \(\beta\), camera-to-mirror distance \(b_m\), and mirror height \(h_m=l_m\sin\beta\). For a single virtual camera, the field of view is
\[
\alpha_{\mathrm{virtual}}=\tan^{-1}\!\Bigl(\frac{l_m\cos\beta}{2b_m + l_m\sin\beta}\Bigr)+\tan^{-1}\!\Bigl(\frac{l_m\cos\beta}{2b_m - l_m\sin\beta}\Bigr).
\]
For \(\beta\approx 55^\circ\), \(b_m=2.5\,\mathrm{cm}\), and \(l_m=3\,\mathrm{cm}\), the paper reports \(\alpha_{\mathrm{virtual}}\approx 47.1^\circ\), which is approximately \(59\%\) of the stock single-camera field of view \((\alpha_{\mathrm{real}}\approx 80^\circ)\) [2109.11872].

Because the virtual cameras tilt toward each other, the inner convergence angle is
\[
\alpha_{in}=90^\circ+\alpha_R-2\beta\approx 14.1^\circ.
\]
For a full \(1.8\,\mathrm{m}\)-tall person, the minimal camera-object distance is reported as
\[
d_{\min}=\frac{b + h_{\mathrm{avg}}}{2\tan\alpha_{in}}\approx 3.69\,\mathrm{m},
\]
with \(b\approx 5\,\mathrm{cm}\) and \(h_{\mathrm{avg}}=1.8\,\mathrm{m}\). Only about \(29.4\%\) of the original vertical field of view is shared by both virtual cameras [2109.11872]. Mechanically, the mask is printed in PLA, includes adjustable slots for mirror angle and tuning for different phone models, and uses a screw-driven slider to vary \(b_m\). The mirrors sit in \(0.5\,\mathrm{mm}\)-precision grooves.

Calibration uses Zhang’s chessboard method in MATLAB. Each raw mirror image is flipped and rotated so that both virtual images share the same upright convention and optical axis direction. The calibration procedure uses 14 stereo pairs of a \(9\times 6\) chessboard with \(2\,\mathrm{cm}\) squares, with autofocus and auto-exposure locked [2109.11872]. The virtual stereo transform \(X_2=RX_1+t\) yields an estimated baseline \(\|t\|\approx 5.4\,\mathrm{cm}\), matching the design. Reprojection error is minimized with nonlinear Levenberg-Marquardt, and the final mean reprojection error is approximately \(1.23\) pixels. Distances between reconstructed 3D points on the calibration planes and their true plane positions are on average \(3\)–\(4\,\mathrm{mm}\) along \(X,Y,Z\) [2109.11872].

The reconstruction pipeline runs each virtual image through OpenPose, producing up to 25 human-joint detections. Since the keypoints are already semantically matched across views, no additional stereo matching is required. Triangulation uses either the linear system
\[
x_{ik}\times(P_iX_k)=0
\]
or the horizontal-stereo depth-from-disparity form
\[
z_k=\frac{b\,f}{x_{1k}-x_{2k}},\qquad
X_k=z_k\,K_1^{-1}[x_{1k},y_{1k},1]^\top.
\]
On a standing subject at approximately \(4\,\mathrm{m}\), six anthropometric segments were evaluated against manual tape measurements. The reconstructed lengths were \([21.3, 24.1, 26.1, 21.3, 38.3, 48.6]\) cm, manual lengths were \([21.2, 24.7, 31.0, 22.1, 38.9, 43.5]\) cm, and the corresponding errors were \([0.1, -0.6, -4.9, -1.2, -0.6, 5.1]\) cm, with mean absolute error approximately \(2.1\,\mathrm{cm}\) [2109.11872].

The paper attributes the largest deviations to keypoint mis-detections at elbows and knees and to the relatively small baseline of approximately \(5\,\mathrm{cm}\), which causes depth sensitivity issues. Proposed improvements include larger mirrors to increase baseline, asymmetric mirror tilts or variable-focus auxiliary optics to enlarge stereo overlap, more varied chessboard views and sub-pixel corner detection for calibration, and fusion with model-based pose priors or temporal filtering for more stable 3D skeletons [2109.11872].

## 3. StereoAdapter as online and federated adaptation for deep stereo

In “Learning to Adapt for Stereo,” StereoAdapter denotes a “learning-to-adapt” framework for robust stereo matching under environmental variation [1904.02957]. The base network is DispNet-Corr1D, written as
\[
d=f_\theta(I_L,I_R),
\]
with \(\theta\) the network parameters. The central objective is to learn an initialization \(\theta\) that is especially amenable to rapid unsupervised online adaptation.

The meta-learning setup uses short stereo sequences \(V^\tau=[i^\tau_1,\dots,i^\tau_n]\). For each sequence, \(k\) consecutive inner-loop steps minimize an unsupervised loss \(\mathcal L_u\), while the outer loop evaluates a supervised loss \(\mathcal L_s\) on the next frame:
\[
\theta^\tau_t=\theta^\tau_{t-1}-\alpha\nabla_\theta \mathcal L_u(\theta^\tau_{t-1}, i^\tau_t),\qquad \theta^\tau_0=\theta.
\]
The outer-loop objective minimizes
\[
\sum_{V^\tau\in\mathcal D_s}\sum_{t=1}^k \mathcal L_s(\theta^\tau_t, i^\tau_{t+1}).
\]
The unsupervised loss combines photometric reprojection, edge-aware smoothness, and left-right consistency:
\[
\mathcal L_u=\lambda_{\rm photo}\mathcal L_{\rm photo}+\lambda_{\rm smooth}\mathcal L_{\rm smooth}+\lambda_{\rm lr}\mathcal L_{\rm lr}.
\]
The photometric term uses SSIM and \(L_1\), smoothness is edge-aware, and left-right consistency uses predictions from both view orderings [1904.02957].

A learned confidence measure is then appended. A small 3-layer fully convolutional network \(\mathcal F_\eta\) takes the pixel-wise error map \(\varepsilon\) derived from \(\mathcal L_{\rm photo}+\mathcal L_{\rm lr}\) and outputs a confidence mask \(W=\mathcal F_\eta(\varepsilon)\in [0,1]^{H\times W}\), normalized so that \(\sum_p W(p)=1\). The weighted unsupervised loss becomes
\[
\mathcal L_u^{\rm masked}=\sum_p W(p)\,\varepsilon(p)+\lambda_{\rm smooth}\mathcal L_{\rm smooth}.
\]
At test time, no ground truth is used. The model updates online from each incoming stereo pair by computing \(W_t\) and taking a gradient step with respect to the weighted unsupervised objective [1904.02957].

The reported training recipe includes supervised pretraining on FlyingThings3D, meta-training on synthetic sequences such as Carla or Synthia with \(\alpha=10^{-5}\), \(\beta=10^{-4}\), \(k=3\), and online test-time adaptation with momentum \(0.9\) and learning rate \(\alpha=10^{-4}\) [1904.02957]. On synthetic-to-real transfer, the best reported variant, L2A + weighted adaptation, achieves D1-all approximately \(4.49\%\) and EPE approximately \(1.12\), compared with \(5.26\%\) and \(1.20\) for supervised learning plus unsupervised online adaptation and \(7.46\%\) and \(1.48\) for static supervised learning on Carla \(\rightarrow\) KITTI. The paper states that this matches the “ideal” fine-tuned model at \(4.26\%\) D1-all and \(1.12\) EPE [1904.02957].

This adaptation-oriented interpretation of StereoAdapter is extended in “Federated Online Adaptation for Deep Stereo,” which introduces a distributed framework based on federated learning [2405.14873]. There, a pre-trained stereo model with parameters \(w\in\mathbb R^P\) is adapted collaboratively across clients observing non-stationary stereo streams \(b_t=(I_t^l, I_t^r)\). The active set \(A\) performs local adaptation, while the listening set \(C\) receives aggregated weights without adapting locally. The server uses FedAvg:
\[
\bar w=\frac{1}{|A|}\sum_{k\in A} w^k.
\]

The underlying network is MADNet 2, an encoder-decoder stack of five scales that replaces original radius-\(r\) correlation volumes with all-pairs correlation from RAFT-Stereo, samples a fixed number \(C\) of channels at each scale, and removes the heavy context network to keep runtime low [2405.14873]. This supports both full adaptation and modular adaptation, the latter updating only one decoder block per step.

The paper reports, on KITTI single-device adaptation, that MADNet 2 without adaptation reaches D1-all \(4.21\) and EPE \(1.09\), while FULL (photo) adaptation reaches D1-all \(1.25\), EPE \(0.83\), and MAD (photo) reaches D1-all \(1.26\), EPE \(0.84\). FULL++ (proxy) and MAD++ (proxy) obtain D1-all \(1.06\), EPE \(0.82\), and D1-all \(1.16\), EPE \(0.84\), respectively [2405.14873]. In the federated setting with three active clients and update interval \(T=10\), FedFULL yields on KITTI D1-all values \(1.42, 1.22, 3.93, 1.12\) across City, Residential, Campus\(\times 2\), and Road, whereas FedMAD gives \(1.48, 1.29, 4.05, 1.16\). Communication is \(20.2\,\mathrm{MB/s}\) to server and \(6.6\,\mathrm{MB/s}\) to client for FedFULL, versus \(4.3\,\mathrm{MB/s}\) and \(3.6\,\mathrm{MB/s}\) for FedMAD [2405.14873].

These results frame StereoAdapter not simply as a model family, but as a deployment paradigm in which stereo systems are designed for unsupervised adaptation during operation, including collaborative adaptation across devices.

## 4. Geometry-aware adaptation and cross-view attention

A further development appears in “Geometry-Aware Unsupervised Domain Adaptation for Stereo Matching,” summarized in the provided data as another StereoAdapter line [2103.14333]. The problem addressed is that image-to-image translation applied independently to the left and right views of a stereo pair can destroy epipolar geometry, leading to mismatched translated pairs and degraded downstream disparity learning.

The proposed mechanism is Stereoscopic Cross Attention (SCA), inserted into a translation network \(\mathcal G\) based on TSIT. At a given scale, each view has a feature map \(F^v\in\mathbb R^{d_{\rm in}\times H\times W}\) and a reprojected 3D point cloud \(P^v\in\mathbb R^{3\times H\times W}\). Query and key projections are computed from the concatenated representation \([F^v;P^v]\), while the value is \(F^v\). Cross-attention is constrained to the horizontal epipolar line over disparity candidates \(d\in [0,d_{\max}]\):
\[
A_{ij,d}^{\,l\to r}
=
\frac{
\exp\!\bigl\langle Q^l_{ij},K^r_{i-d,j}\bigr\rangle/\sqrt{d_{\rm out}}
}{
\sum_{d'=0}^{d_{\max}}
\exp\!\bigl\langle Q^l_{ij},K^r_{i-d',j}\bigr\rangle/\sqrt{d_{\rm out}}
},
\]
\[
\mathrm{SCA}_{ij}^{l\to r}
=
\sum_{d=0}^{d_{\max}}
A_{ij,d}^{\,l\to r}\,V^r_{i-d,j}.
\]
The same construction is applied in the reverse direction [2103.14333].

The overall pipeline has three stages: source-supervised pretraining of a stereo matcher \(\mathcal E\), pixel-level adaptation through \(\mathcal G\) and a discriminator \(\mathcal C\), and target training or fine-tuning of \(\mathcal E\) on translated pairs plus optional unsupervised reprojection on target stereo pairs [2103.14333]. The losses include adversarial, perceptual, feature matching, stereo-consistency, supervised disparity, and target reprojection terms. The stereo-consistency loss is computed on translated multi-scale features using known source disparities and an occlusion mask.

Quantitatively, on Driving \(\rightarrow\) KITTI 2015 with a DispNetC backbone, the paper reports D1-all \(10.11\%\), EPE \(1.64\) px without SCA, and D1-all \(8.40\%\), EPE \(1.47\) px with SCA [2103.14333]. In comparison to other unsupervised methods, the proposed approach reaches D1-all \(8.40\%\), EPE \(1.47\) px, compared with StereoGAN at \(25.71\%\), \(2.75\) px and L2A+Wad at \(26.90\%\), \(3.01\) px. With a GwcNet backbone on SceneFlow \(\rightarrow\) KITTI, the proposed method reaches D1-all \(5.19\%\), EPE \(1.12\) px [2103.14333].

This line of work is methodologically related to later “adapter” approaches in that it introduces a stereo-specific cross-view mechanism into a pre-existing architecture. The difference is that the inserted mechanism is geometry-preserving attention for domain translation, not parameter-efficient fine-tuning in the modern PEFT sense.

## 5. Adapter modules for stereo image super-resolution

In “ASTeISR: Adapting Single Image Super-resolution Pre-trained Model for Efficient Stereo Image Super-resolution,” the phrase “stereo adapter” has a more specific PEFT meaning [2407.03598]. The method starts from a frozen, pre-trained single-image super-resolution transformer, HAT-L, and injects two kinds of lightweight modules: stereo adapters for cross-view fusion and spatial adapters for local feature refinement.

Stereo adapters are placed immediately after each Residual Hybrid Attention Group (RHAG), while spatial adapters are inserted inside each Hybrid Attention Block (HAB), after the hybrid-attention layer and before the feed-forward MLP [2407.03598]. For a pair of feature tensors \(X_L, X_R\in\mathbb R^{H\times W\times C}\), the stereo adapter applies LayerNorm, linear projections, and temperature-scaled cross-attention:
\[
F_{R\to L}=TA(Q=\bar X_LW_1^L,\;K=\bar X_RW_1^R,\;V=X_RW_2^R),
\]
\[
TA(Q,K,V)=\mathrm{softmax}(\tau\cdot QK^\top/\sqrt C)V.
\]
The residual updates are
\[
\hat X_L=\gamma_L\cdot F_{R\to L}+X_L,\qquad
\hat X_R=\gamma_R\cdot F_{L\to R}+X_R.
\]
The spatial adapter is a bottleneck MLP within each HAB:
\[
f_{\rm spa}(H)=W_2\cdot GELU(W_1\cdot H),\qquad
\hat H=H+f_{\rm spa}(H).
\]
The reduction ratio is \(r=16\) in practice [2407.03598].

Fine-tuning keeps the base HAT-L model, approximately \(40.7\)M parameters, fully frozen. Only the stereo adapters and spatial adapters are trained, totaling approximately \(1.97\)M tunable parameters, or approximately \(4.8\%\) of the base [2407.03598]. The loss is \(L_1\) between super-resolved left and right outputs and their ground-truth high-resolution targets. Training uses AdamW with initial learning rate \(5\times 10^{-4}\), no extra regularizer beyond weight decay, Flickr1024 plus Middlebury with random \(128\times 320\) crops, \(2\times\) A40 GPUs, batch size \(3\), and \(200\)K iterations [2407.03598].

The reported quantitative gains center on Flickr1024. For \(4\times\) super-resolution, frozen HAT-L yields \(29.54\) dB, full fine-tuning yields \(29.53\) dB, PEFT with spatial adapters only gives \(29.55\) dB, stereo adapters only give \(29.66\) dB, and stereo plus spatial adapters give \(29.74\) dB [2407.03598]. For \(2\times\) super-resolution on Flickr1024, HAT-L gives \(29.54\) dB and ASteISR gives \(30.33\) dB, a \(+0.79\) dB gain. Compared with full fine-tuning, the method reduces fine-tuning time from \(127\) h to \(55\) h and GPU memory from \(25.1\) GB to \(21.4\) GB, corresponding to reductions of \(57\%\) and \(15\%\), respectively [2407.03598].

Ablations indicate that the spatial adapter alone provides negligible gain, approximately \(+0.01\) dB, whereas the stereo adapter alone provides \(+0.12\) dB, and both together provide \(+0.20\) dB in the \(4\times\) setting. Performance increases as more stereo-adapter layers are added and saturates after approximately \(12\)–\(16\) adapters, one per RHAG. Alternative spatial-adapter placements in parallel to the MLP or after the MLP increase tunable parameters and GPU time by at least \(2\times\) but slightly hurt PSNR by approximately \(-0.04\) to \(-0.08\) dB [2407.03598]. The paper also states that full fine-tuning on the small stereo dataset degrades texture restoration relative to the frozen SISR prior, whereas stereo adapters recover and improve on that prior.

This work is representative of a newer use of “adapter” in the PEFT literature: instead of adapting a stereo model online, the method adapts a monocular pre-trained model into a stereo-capable one by adding small trainable modules.

## 6. Underwater StereoAdapter and StereoAdapter-2

In “StereoAdapter: Adapting Stereo Depth Estimation to Underwater Scenes,” the term denotes a self-supervised underwater stereo depth estimation framework that addresses severe underwater domain shift, including wavelength-dependent attenuation, backscattering and forward scattering, refraction at water-glass interfaces, specular highlights, and textureless regions [2509.16415]. The paper identifies two key challenges: adapting large pre-trained vision encoders to underwater appearance in a parameter-efficient manner without dense depth labels, and fusing globally coherent but scale-ambiguous monocular priors with locally metric yet photometrically fragile stereo correspondences.

The architecture is two-stage. Stage 1 uses a LoRA-adapted Depth Anything V2 encoder, frozen except for inserted LoRA modules, to generate multi-scale features
\[
\{F_i\in\mathbb R^{C\times(H/2^i)\times(W/2^i)}\}_{i=2}^5,
\]
aggregate them with SDFA blocks into a discrete disparity volume \(V_m\in\mathbb R^{N\times H\times W}\), and produce a coarse disparity map \(d^{(1)}\) and metric depth \(M^{\rm mono}\) [2509.16415]. The LoRA update for each transformer weight \(W_0\in\mathbb R^{d\times k}\) is
\[
\Delta W=BA,\qquad B\in\mathbb R^{d\times r},\ A\in\mathbb R^{r\times k},\ r\ll \min(d,k).
\]

Stage 2 performs recurrent stereo refinement. Stereo feature maps \(f_L,f_R\in\mathbb R^{C\times H/4\times W/4}\) are used to build a 4D correlation volume
\[
C(i,j,d)=\langle f_L(i,j),f_R(i,j-d)\rangle,\qquad d\in [0,D_{\max}],
\]
followed by hybrid scale alignment to verify or correct the global scale of \(M^{\rm mono}\) through sparse stereo correspondences \(\{D_{\rm sparse}(p)\}\), multi-scale context aggregation, and iterative ConvGRU refinement:
\[
h^{(l+1)},\Delta d^{(l)}={\rm ConvGRU}\bigl(h^{(l)},[c^{(l)},g^{(l)},\phi(d^{(l)})]\bigr).
\]
The final disparity is upsampled RAFT-style to full resolution [2509.16415].

The method also introduces dynamic LoRA adaptation, in which the update is written as
\[
\Delta W=\sum_{i=1}^{r_{\max}} w_i B_iA_i,
\]
with \(\ell_1\) sparsity on the importance weights \(w_i\). The joint training objective is
\[
\mathcal L_{\rm train}=\mathcal L_{\rm sup}+\lambda\sum_{m=1}^M \|w^m\|_1,
\]
and the importance weights are updated using a proximal-gradient soft-threshold rule with a threshold \(\kappa\) ramped from \(0\) to \(\kappa_{\max}\) [2509.16415].

Self-supervised learning is divided into monocular and stereo losses. The monocular stage uses
\[
\mathcal L_{\rm mono}=\mathcal L_{\rm rec}^{\rm mono}+\lambda_1 \mathcal L_{\rm smooth}^{\rm mono},
\]
while the stereo stage uses
\[
\mathcal L_{\rm stereo}
=
\mathcal L_{\rm rec}^{\rm stereo}
+\lambda_3\mathcal L_{\rm smooth}^{\rm stereo}
+\lambda_4\mathcal L_{\rm guide}.
\]
The guidance term constrains agreement between the coarse monocular disparity \(d^{(1)}\) and the refined stereo disparity \(d^{(L)}\) in both gradients and values, including an outlier mask \(\mathbf M_{\rm out}\) [2509.16415].

Pre-training uses the synthetic UW-StereoDepth-40K dataset rendered in Unreal Engine 5 with caustics, scattering, color attenuation, and particles. The four scene types are coral reefs, shipwrecks, industrial structures, and natural seabed. Baselines are \(\{4\,\mathrm{cm},10\,\mathrm{cm},20\,\mathrm{cm},40\,\mathrm{cm}\}\), and the dataset contains \(40{,}000\) stereo pairs at \(1280\times 960\) with dense ground-truth depth and segmentation [2509.16415].

The reported results cover zero-shot, fine-tuned, and real-world deployment. On TartanAir Underwater, zero-shot StereoAdapter trained on UW-StereoDepth-40K yields REL \(0.0527\), SQ REL \(0.5167\), RMSE \(2.8947\), and \(A1\) \(94.67\%\), compared with Stereo Anywhere at REL \(0.0592\), RMSE \(3.1572\), and FoundationStereo at REL \(0.0542\), RMSE \(2.9644\). Fine-tuning on TartanAir gives REL \(0.0512\), RMSE \(2.7834\), \(A1\) \(95.12\%\) [2509.16415]. On SQUID, zero-shot StereoAdapter yields REL \(0.0806\), RMSE \(1.8843\), \(A1\) \(94.13\%\), and the fine-tuned version yields REL \(0.0795\), RMSE \(1.8621\), \(A1\) \(94.28\%\). On BlueROV2 indoor tank deployment, StereoAdapter yields REL \(0.0856\), SQ REL \(0.6482\), RMSE \(1.9690\), and \(A1\) \(94.78\%\). Test-time latency on Jetson Orin NX at \(640\times 360\), batch size \(1\), is \(1113\) ms, compared with \(1815\) ms for FoundationStereo and \(1440\) ms for Stereo Anywhere [2509.16415].

“StereoAdapter-2” revises this design by retaining the LoRA-adapted monocular foundation encoder, now Depth Anything 3 ViT-B, for feature extraction and disparity initialization, while replacing the ConvGRU-based updater with a ConvSS2D operator based on selective state-space models [2602.16915]. The continuous-time SSM
\[
h'(t)=A\,h(t)+B\,u(t),\qquad y(t)=C\,h(t)+D\,u(t)
\]
is discretized to a structured convolution kernel and 1D recurrence
\[
h_t=\bar A h_{t-1}+\bar B x_t.
\]
In ConvSS2D, the recurrence is input-dependent:
\[
h_t=\bar A_t h_{t-1}+\bar B_t x_t,
\]
with
\[
\Delta_t=\mathrm{softplus}(W_\Delta x_t),\qquad
\bar B_t=W_Bx_t,\qquad
C_t=W_Cx_t.
\]
The scan is applied in four directions: left-to-right, right-to-left, top-to-bottom, and bottom-to-top [2602.16915]. The horizontal directions align with epipolar geometry, while the vertical directions enforce structural consistency.

The complexity discussion contrasts ConvGRU and ConvSS2D. ConvGRU requires approximately four convolutions of kernel size \(k\times k\) over \(H\times W\), giving \(O(I\cdot H\cdot W\cdot k^2\cdot C^2)\), whereas ConvSS2D performs four linear recurrences of length \(H\cdot W\) with \(O(d_{\rm state}^2)\) per location, giving \(O(H\cdot W\cdot d_{\rm state}^2\cdot 4)\). Since \(d_{\rm state}\) is small, such as \(4\), the method is effectively linear in the number of pixels with a smaller constant [2602.16915].

StereoAdapter-2 also expands the synthetic corpus to UW-StereoDepth-80K by generating \(40\)K new diffusion-based samples and merging them with UW-StereoDepth-40K. The pipeline combines Atlantis style transfer conditioned on source depth with NVS-Solver novel-view synthesis. Baselines are \(\{20,30,40,50\}\) cm, and the resulting stereo pairs are \(640\times 480\) with ground-truth disparity [2602.16915].

The zero-shot TartanAir-UW results reported for StereoAdapter-2 trained on \(80\)K are AbsRel \(0.0440\), SqRel \(0.4312\), RMSE \(2.4038\), LogRMSE \(0.1198\), \(\delta_1=96.76\%\), \(\delta_2=97.04\%\), and \(\delta_3=98.90\%\). Compared with StereoAdapter-1 trained on \(40\)K, which gives AbsRel \(0.0527\) and RMSE \(2.8947\), this corresponds to a \(16.5\%\) REL reduction and \(17.0\%\) RMSE reduction [2602.16915]. On SQUID, StereoAdapter-2 gives AbsRel \(0.0705\), RMSE \(1.7481\), and \(\delta_1=94.25\%\), improving over StereoAdapter-1’s AbsRel \(0.0806\) and RMSE \(1.8843\). On BlueROV2, it reports REL \(0.1023\), SqRel \(0.5843\), RMSE \(1.7164\), LogRMSE \(0.1354\), and \(\delta_1=92.56\%\), outperforming the cited baseline FoundationStereo at REL \(0.1304\), RMSE \(2.0893\) [2602.16915]. On-board inference with Jetson Orin NX and TensorRT at \(640\times 360\), batch \(1\), is \(1.102\,\mathrm{s/frame}\) [2602.16915].

## 7. Common themes, distinctions, and interpretive context

Across these works, “StereoAdapter” consistently refers to a mechanism for making stereo systems usable outside the narrow regime assumed by a base setup or model. In the smartphone rig, the adaptation is optical: planar mirrors adapt monocularly separated front and rear cameras into a stereo pair [2109.11872]. In online and federated stereo matching, the adaptation is parametric: model weights are updated to accommodate domain shift during deployment [1904.02957], [2405.14873]. In geometry-aware domain adaptation, the adaptation is representational: cross-view attention preserves epipolar structure during image translation [2103.14333]. In ASteISR, adaptation is parameter-efficient transfer from single-image super-resolution to stereo super-resolution using inserted modules [2407.03598]. In underwater StereoAdapter and StereoAdapter-2, adaptation combines PEFT, self-supervision, and recurrent stereo refinement to bridge the severe physics-induced gap between terrestrial and underwater imagery [2509.16415], [2602.16915].

A common misconception would be to treat these as successive versions of a single architecture. The corpus does not support that interpretation. Only the underwater pair, StereoAdapter and StereoAdapter-2, forms an explicit generational sequence [2509.16415], [2602.16915]. The smartphone device, learning-to-adapt framework, federated adaptation system, geometry-aware SCA method, and ASteISR are independent uses of the same or closely related naming convention.

Another distinction concerns what is being “adapted.” In the 2019 and 2024 online-adaptation works, a stereo matcher is adapted to the environment [1904.02957], [2405.14873]. In the 2021 geometry-aware method, the source images are adapted to the target domain while preserving stereo geometry [2103.14333]. In ASteISR, a single-image pre-trained transformer is adapted to stereo super-resolution [2407.03598]. In the 2025 and 2026 underwater works, a monocular foundation encoder is adapted to underwater stereo depth estimation, and then the stereo updater itself is redesigned for more global propagation [2509.16415], [2602.16915]. In the smartphone rig, the cameras are physically adapted into a stereo apparatus [2109.11872].

This suggests that “StereoAdapter” has become a productive research label precisely because it is semantically broad: it can denote a device, an online learning rule, a federated adaptation framework, an attention mechanism, a PEFT module, or a domain-specific stereo depth system. The unifying idea is stereo functionality obtained through an added structure—optical, architectural, algorithmic, or optimization-based—rather than through a stereo system designed from scratch.

Source: https://www.emergentmind.com/topics/stereoadapter