MedBayes-Lite: Efficient Stereo Matching

• MedBayes-Lite is an ambiguous label for a family of lightweight stereo matchers that emphasize efficient cost-volume processing, low latency, and robust zero-shot transfer.
• These methods replace heavy 4D cost volumes with quarter-resolution correlation and 2D aggregation or minimal 3D refinements, streamlining dense disparity estimation.
• They employ varied training strategies—from synthetic pretraining to cost volume stabilization—to ensure accurate, real-world performance under strict computational constraints.

Searching arXiv for “MedBayes-Lite” and closely related names to verify whether a paper with that title exists. Searching arXiv for “MedBayes-Lite”. MedBayes-Lite is not the name of a method in the lightweight stereo-matching literature represented by "LightStereo" (Guo et al., 2024), "LeanStereo" (Rahim et al., 24 Mar 2025), "Lite Any Stereo" (Jing et al., 20 Nov 2025), "Lite Any Stereo V2" (Jing et al., 23 Jun 2026), and "LiteMatch" (Khan et al., 30 Jun 2026). The term is therefore best treated as an ambiguous label whose nearest identifiable referents are these efficient stereo systems. Taken together, those works define a coherent research direction: dense disparity estimation from rectified stereo pairs under strict latency, memory, and deployment constraints, with particular emphasis on replacing heavy 4D cost volumes, reducing 3D CNN usage, and improving zero-shot transfer across real-world domains.

1. Terminological status and nearest identifiable lineage

The nearest published lineage to the label MedBayes-Lite is a family of lightweight stereo matchers rather than a single canonical method. "LightStereo" studies efficient 2D cost aggregation over a correlation-based 3D cost volume and explicitly argues that channel or disparity mixing is the dominant lever for recovering accuracy lost by abandoning heavier 4D cost processing (Guo et al., 2024). "LeanStereo" remains within the 3D cost-volume paradigm but replaces the backbone with a much leaner two-branch design and compensates with an attention-weighted cost volume and LogL1 loss (Rahim et al., 24 Mar 2025). "Lite Any Stereo" introduces a hybrid 3D-2D aggregation module together with a three-stage million-scale training strategy for efficient zero-shot stereo matching (Jing et al., 20 Nov 2025). "Lite Any Stereo V2" then replaces the hybrid aggregation with a deployment-oriented 2D-only framework and extends the training pipeline with pseudo-label filtering and error clamping (Jing et al., 23 Jun 2026). "LiteMatch" approaches the same efficiency–generalization problem through cost-volume stabilization, dual encoders, and CVC-Loss rather than through large-scale real-data distillation alone (Khan et al., 30 Jun 2026).

A frequent source of confusion is that these systems are related in naming but not identical in design philosophy. LightStereo is explicitly not an anytime stereo method in the sense of AnyNet; its contribution is lightweight correlation-volume aggregation rather than progressive inference (Guo et al., 2024). Lite Any Stereo and Lite Any Stereo V2 are zero-shot stereo systems organized around staged sim-to-real training, whereas LeanStereo is a lightweight 3D stereo network optimized primarily for speed and operation count, and LiteMatch frames the central problem as stabilizing the cost volume itself rather than repairing it later with heavier regularization (Rahim et al., 24 Mar 2025, Jing et al., 20 Nov 2025, Khan et al., 30 Jun 2026).

2. Problem formulation and shared stereo pipeline

Across these works, stereo matching is the standard dense correspondence problem on rectified left/right RGB images. The objective is to estimate a disparity field, from which depth can be recovered. The shared computational bottleneck is cost-volume processing: many classical deep stereo systems construct a feature-rich 4D cost volume over disparity and spatial dimensions and regularize it with 3D CNNs, which drives up FLOPs, memory traffic, and latency, particularly as disparity range grows (Guo et al., 2024).

The lightweight line instead tends to construct quarter-resolution matching features and use correlation-style costs. In LightStereo, the quarter-scale left and right features $\mathbf f_{l,4}$ and $\mathbf f_{r,4}$ are used to form a correlation volume

$$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$

yielding a tensor with shape roughly $(D/4,H/4,W/4)$ that is then processed as disparity-as-channels by a 2D encoder-decoder (Guo et al., 2024). Lite Any Stereo uses the same general correlation formulation at quarter resolution,

$$\mathbf{C}(d, h, w) = \frac{1}{N_c} \left\langle \mathbf{F}_{L}^{1/4}(h, w),\, \mathbf{F}_{R}^{1/4}(h, w-d) \right\rangle,$$

followed by hybrid aggregation and soft disparity regression (Jing et al., 20 Nov 2025). LAS2 preserves the quarter-resolution correlation-volume structure, fixes $D_{\max}=192$, and defines its feed-forward path as feature extraction, correlation, 2D-only aggregation, soft argmax, and convex upsampling (Jing et al., 23 Jun 2026).

LiteMatch uses the same broad quarter-scale logic but emphasizes that the instability of the early cost distribution is the real source of downstream complexity. Its notation specifies a cost tensor $C \in \mathbb{R}^{H \times W \times D}$ and a disparity probability volume

$P_{i,j,d} = \mathrm{Softmax}_d(C_{i,j,d}),$

which becomes the target of direct stabilization by CVC-Loss (Khan et al., 30 Jun 2026). This suggests that the most important conceptual divide in the literature is no longer simply 2D versus 3D aggregation, but where the method chooses to spend its limited capacity: backbone design, disparity-channel mixing, tiny 3D injections, or training-time stabilization of the probability volume.

3. Architectural patterns in efficient stereo matching

The identifiable methods near the MedBayes-Lite label span several distinct architectural strategies.

Method	Core aggregation idea	Main deployment claim
LightStereo	2D encoder-decoder on a correlation-based 3D cost volume, with channel boost	Speed, accuracy, and resource utilization
LeanStereo	Lean backbone plus attention-weighted cost volume and 3D hourglass regularization	4× fewer operations and about 9× to 14× faster inference than ACVNet, LEAStereo, and CFNet
Lite Any Stereo	Hybrid 3D-2D serial aggregation	Strong zero-shot generalization at 33 G MACs
Lite Any Stereo V2	Purely 2D cost aggregation with FasterNet-style modules	Real latency rather than theoretical MACs
LiteMatch	Dual encoder plus 2D cost refinement and stabilized probability volume	Strong zero-shot generalization without expensive 3D convolutions

LightStereo’s architectural thesis is that a 3D correlation volume already encodes the crucial matching-cost distribution along the disparity or channel dimension, so the main job of aggregation is dense channel mixing rather than expansive spatial reasoning. Its MobileNetV2-style inverted residual blocks therefore expand channels with a $1\times1$ convolution, apply a lightweight depthwise $3\times3$ spatial filter, and project back with another $\mathbf f_{r,4}$0 convolution, using ReLU6 and residual connections when dimensions match (Guo et al., 2024). The same paper adds MSCA, a Multi-Scale Convolutional Attention module that processes left-image features at $\mathbf f_{r,4}$1, $\mathbf f_{r,4}$2, and $\mathbf f_{r,4}$3 scales with strip-like depthwise separable convolutions using kernels $\mathbf f_{r,4}$4, $\mathbf f_{r,4}$5, $\mathbf f_{r,4}$6, $\mathbf f_{r,4}$7, $\mathbf f_{r,4}$8, $\mathbf f_{r,4}$9, and $$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$0, then gates the aggregated cost (Guo et al., 2024).

LeanStereo takes the opposite stance on cost regularization. It preserves a full 3D regularization network with four 3D convolution layers followed by two stacked hourglass encoder-decoders, but shifts most of the efficiency work into the feature extractor. The backbone is split into a shallow high-resolution branch and a deeper low-channel semantic branch inspired by MobileNetV2, PeleeNet, and BiSeNetV2, with guided aggregation between them. The cost volume is neither pure concatenation nor pure correlation: a concatenation-based pre-attention volume is modulated by attention weights inferred from a group-wise correlation volume (Rahim et al., 24 Mar 2025).

Lite Any Stereo occupies a middle position. It uses a compact MobileNetV2-style backbone, quarter-resolution correlation, and a hybrid aggregation module defined as

$$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$1

where the 3D block uses multi-scale 3D convolutions to preserve disparity perception and the 2D block uses ConvNeXt layers to enhance spatial perception (Jing et al., 20 Nov 2025). Its ablations show that the best arrangement is 3D-then-2D serial composition, that standard $$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$2 3D kernels work best, and that only a very small 3D fraction is beneficial: the best setting uses 4.8% 3D proportion (Jing et al., 20 Nov 2025).

LAS2 then revises this design under the criterion of measured deployment latency. It removes the 3D portion entirely from the feed-forward path and replaces it with a purely 2D U-Net-style encoder-decoder over the cost representation, implemented with FasterNet-style residual blocks and the attention mechanism retained from LightStereo (Jing et al., 23 Jun 2026). The paper is explicit that the shift is motivated not only by lower compute but by lower real latency on hardware such as H200 and Orin NX 8G. LAS2-H adds recurrent refinement on top of LAS2-M, but its initialization still comes from the efficient 2D aggregation front-end rather than from a heavy 3D geometry encoder (Jing et al., 23 Jun 2026).

LiteMatch reorganizes the architecture around feature quality and probability-volume stability. It uses two complementary encoders: a Cross-View Correspondence Encoder that applies bidirectional cross-attention before cost-volume construction, and a High-Frequency Encoder that preserves fine structure by FFT-based high-pass filtering,

$$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$3

with low-frequency components removed in the central frequency region. The fused features are used to build the cost volume, which is then refined by lightweight 2D feature-guided residual updates rather than by 3D regularization (Khan et al., 30 Jun 2026). A plausible implication is that LiteMatch treats correspondence formation itself as the primary locus of efficiency, whereas LightStereo and LAS/LAS2 place more emphasis on the aggregation path after correlation.

4. Training regimes and zero-shot generalization

The methods closest to MedBayes-Lite differ most sharply in training strategy. LightStereo uses a comparatively simple supervised recipe: smooth $$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$4 disparity loss, AdamW, OneCycleLR, random crop on Scene Flow, and fine-tuning on mixed KITTI 2012+2015 with color jitter, random erase, random scale, and random crop (Guo et al., 2024). Its contribution lies mainly in architecture and ablation, not in synthetic-to-real adaptation machinery.

LeanStereo is also relatively conventional in training, using Adam, $$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$5 crops, maximum disparity $$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$6, ImageNet normalization, 900K Scene Flow iterations, and KITTI fine-tuning from the synthetic checkpoint. Its main supervisory novelty is LogL1 loss,

$$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$7

introduced to model and improve smaller disparity errors more explicitly and reported to yield better EPE, better D1, improved bad-pixel rates, faster convergence, and more stable training (Rahim et al., 24 Mar 2025).

Lite Any Stereo and LAS2 instead treat training as coequal with architecture. Both use a three-stage strategy: supervised synthetic training, self-distillation on synthetic data, and knowledge distillation on unlabeled real stereo pairs (Jing et al., 20 Nov 2025, Jing et al., 23 Jun 2026). In Lite Any Stereo, the synthetic set totals 1.8M annotated stereo pairs from SceneFlow, FallingThings, FSD, CREStereo, VKITTI2, TartanAir, and Dynamic Replica, while the real stage uses 0.5M unlabeled pairs from Flickr1024, InStereo2K, Holopix50K, DrivingStereo, SouthKenSV, and UASOL (Jing et al., 20 Nov 2025). Stage 2 adds a feature alignment loss,

$$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$8

with a fixed-teacher formulation performing better than EMA or hard-copy teacher updates (Jing et al., 20 Nov 2025). LAS2 preserves this three-stage structure but strengthens Stage 3 with pseudo-label filtering and an error-clamped loss (Jing et al., 23 Jun 2026).

LAS2’s Stage 3 is unusually explicit. The final valid supervision mask combines left-right consistency, edge-aware masking, and a sky mask. The clamped real-data distillation loss is

$$C_{cor}(d, h, w) = \frac{1}{C} \sum_{c=1}^{C} {f}_{l,4}(h, w) \cdot {f}_{r,4}(h, w-d),$$9

with $(D/4,H/4,W/4)$0 in the final model. The motivation is that a few large synthetic-bias errors on real imagery can dominate gradients and destabilize adaptation; the clamped formulation makes Stage 3 behave more like robust regression (Jing et al., 23 Jun 2026). The paper also reports that the same three-stage strategy improves LightStereo-M and BANet-2D, suggesting that the training recipe is not LAS2-specific (Jing et al., 23 Jun 2026).

LiteMatch offers a different training philosophy. It does not use domain adaptation, monocular priors, or extra mixed-dataset training for its zero-shot evaluation. Instead, it trains exclusively on Scene Flow for pretraining and stabilizes the disparity distribution directly with the Cost Volume Consistency Loss,

$(D/4,H/4,W/4)$1

The paper interprets this as standard cross-entropy on the correct disparity bin plus explicit suppression of every incorrect bin, encouraging a sharp and unimodal distribution and lowering entropy from 2.74 to 1.77 relative to CE (Khan et al., 30 Jun 2026). This suggests two competing but compatible explanations for efficient zero-shot stereo: either one must distill real-world robustness into a small model, as in Lite Any Stereo and LAS2, or one must make the cost distribution intrinsically more stable, as in LiteMatch.

5. Efficiency, latency, and benchmark behavior

The common empirical message is not that one method dominates all others, but that several different efficiency strategies can be made competitive.

LightStereo is one of the clearest demonstrations of how much can be gained by replacing 4D cost processing with quarter-resolution correlation and 2D aggregation. On Scene Flow, LightStereo-S uses 22.71 GFLOPs, 3.44M parameters, EPE 0.73, and 17 ms runtime; LightStereo-M uses 36.36 GFLOPs, 7.64M parameters, EPE 0.62, and 23 ms; LightStereo-L uses 91.85 GFLOPs, 24.29M parameters, EPE 0.59, and 37 ms; and LightStereo-L$(D/4,H/4,W/4)$2 uses 159.26 GFLOPs, 45.63M parameters, EPE 0.51, and 54 ms (Guo et al., 2024). On KITTI 2015, LightStereo-L$(D/4,H/4,W/4)$3 reports D1-bg 1.60, D1-fg 2.92, and D1-all 1.82 in 49 ms, and the paper claims it is top on D1-bg and D1-all among the listed lightweight methods (Guo et al., 2024).

LeanStereo represents a strong point in lightweight 3D stereo. On Scene Flow, it reports EPE 0.75, D1 2.47, 66.03 G MACs, and 6.6M parameters, outperforming GwcNet-g on EPE and D1 while using substantially fewer operations (Rahim et al., 24 Mar 2025). On KITTI validation, it reports EPE 0.62, D1 1.78, 60.41 G MACs, and 6.60M parameters, matching the best EPE in the table while trailing GwcNet on D1 (Rahim et al., 24 Mar 2025). Its runtime comparison is particularly sharp: 35 ms on 400 images of size $(D/4,H/4,W/4)$4 for a 16-bit floating TensorRT model, versus 302 ms for ACVNet, 483 ms for CFNet, and 945 ms for LEAStereo (Rahim et al., 24 Mar 2025).

Lite Any Stereo advances the zero-shot argument. At 33 G MACs, it reports million-scale-training results of KITTI 2012 D1 3.04 and EPE 0.79, KITTI 2015 D1 3.87 and EPE 0.99, ETH3D Bad1.0 3.53 and EPE 0.32, and Middlebury Bad2.0 7.51 and EPE 0.94, outperforming LightStereo-M and BANet-2D across the D1/Bad metrics in the efficient-model block while tying them or beating them on almost every EPE metric (Jing et al., 20 Nov 2025). Its hardware timings are also strong: 21 ms on GTX 1080, 19 ms on RTX 4090, 23 ms on A5000, and 17 ms on A100, making it the fastest method in that comparison (Jing et al., 20 Nov 2025).

LAS2 sharpens the deployment case by switching the optimization target from abstract cost to measured latency. At input $(D/4,H/4,W/4)$5, LAS2-S runs in 6.6 ms on H200 and 81 ms on Orin NX 8G MAXN, LAS2-M in 8.1 ms and 101 ms, and LAS2-L in 11.4 ms and 166 ms (Jing et al., 23 Jun 2026). LAS2-H, the iterative variant, runs in 15.1 ms on H200 and 344 ms on Orin, versus 27.3 ms and 918 ms for Fast-FoundationStereo (Jing et al., 23 Jun 2026). On zero-shot benchmarks, LAS2-M reports KITTI 2012 D1 2.88 and EPE 0.74, KITTI 2015 D1 3.61 and EPE 0.95, ETH3D Bad-1.0 2.59 and EPE 0.27, and Middlebury Bad-2.0 5.47 and EPE 0.77 (Jing et al., 23 Jun 2026). LAS2-H reaches KITTI 2012 D1 2.64, KITTI 2015 D1 3.31, ETH3D Bad-1.0 1.83, and Middlebury Bad-2.0 3.71, while remaining substantially faster than the iterative baselines in its table (Jing et al., 23 Jun 2026).

LiteMatch offers a different efficiency profile. The base model uses 3.36M parameters, 79.8 GFLOPs, 1.17 GB memory, and 45 ms runtime at KITTI-sized input $(D/4,H/4,W/4)$6, while the iterative variant uses 9.58M parameters, 179.0 GFLOPs, 1.41 GB memory, and 220 ms (Khan et al., 30 Jun 2026). Its zero-shot results are notable: on DrivingStereo it reports D1-all 2.00% averaged across Sunny, Cloudy, Rainy, and Foggy, with per-condition D1 of 2.01, 2.05, 2.44, and 1.51; on KITTI zero-shot, the base model reports KITTI-2015 EPE 1.21 and Bad3.0 5.40%, KITTI-2012 EPE 1.11 and Bad3.0 5.12%, and 22.2 FPS; and the full iterative model reaches KITTI-2012 Bad3.0 4.20% and KITTI-2015 Bad3.0 4.09% (Khan et al., 30 Jun 2026).

6. Limitations, misconceptions, and broader significance

The nearest literature to MedBayes-Lite is unified by efficiency goals but not by a single settled doctrine. LightStereo shows that 2D aggregation can be highly competitive if channel or disparity mixing is designed correctly, but its own motivation concedes that 2D aggregation historically underperforms heavier 3D methods, and its coarse quarter-resolution correlation volume may be less robust in thin structures, reflective regions, occlusions, and large ambiguous textureless zones (Guo et al., 2024). LeanStereo demonstrates that one can retain a classical 3D cost-volume pipeline while compressing the backbone, but it remains a lightweight 3D stereo model rather than a fully edge-native solution, since it still relies on 4D cost volumes, 3D convolutions, and TensorRT FP16 timing for its headline speed (Rahim et al., 24 Mar 2025).

Lite Any Stereo and LAS2 show that efficient models can be zero-shot-capable, but both papers also delimit the claim. Lite Any Stereo still trails prior-based approaches such as FoundationStereo in absolute accuracy on the hardest zero-shot benchmarks, and it notes that limited high-quality real-world stereo data remains a bottleneck, with Middlebury after Stage 3 likely harmed by the relative scarcity of indoor real data (Jing et al., 20 Nov 2025). LAS2 similarly acknowledges that foundation-model-assisted systems benefit from richer semantic or geometric priors and larger backbones, and it lists remaining failure cases including reflections, transparency, severe illumination change, and ambiguous geometry (Jing et al., 23 Jun 2026).

LiteMatch sharpens a different misconception: that lightweight generalization must come from larger priors or long iterative optimization. Its results suggest that a small stereo-only model can generalize well if the model learns globally coherent, edge-aware features and the cost volume is explicitly regularized to be sharp and unimodal early in training (Khan et al., 30 Jun 2026). Whether that strategy supersedes the three-stage real-data distillation approach remains unresolved. A plausible implication is that the efficient stereo field is moving toward two partially complementary principles: deployment-aware operator selection, emphasized by LAS2, and training-time stabilization of the disparity distribution, emphasized by LiteMatch.

Under that interpretation, MedBayes-Lite is best understood not as a documented standalone system, but as an imprecise label for a rapidly evolving family of efficient stereo matchers. The family’s central contributions are now clear: quarter-resolution correlation volumes, reduced reliance on heavy 3D regularization, selective or minimal use of 3D disparity reasoning when it is retained, deployment-driven latency evaluation, and increasingly sophisticated strategies for cross-domain zero-shot transfer (Guo et al., 2024, Jing et al., 20 Nov 2025, Jing et al., 23 Jun 2026, Khan et al., 30 Jun 2026).