Disparity Has a Sign: Stereo Matching Beyond the Zero-Disparity Plane

Jian Shi; Xinge Yang; Chaoyang Wang; Wolfgang Heidrich; Peter Wonka

cs.CVSep 6, 2026

Disparity Has a Sign: Stereo Matching Beyond the Zero-Disparity Plane

Authors: Jian Shi, Xinge Yang, Chaoyang Wang, Wolfgang Heidrich, Peter Wonka

Organizations: KAUST, Thuwal, Saudi Arabia

Abstract

Modern stereo matching models fail when disparity crosses zero, with end-point error (EPE) rising by 4.6-37 $\times$ . Yet stereoscopic content, from cinema 3D to VR, routinely contains objects behind the zero-disparity plane (ZDP), corresponding to negative disparities. The blind spot cascades through datasets, architectures, and evaluation protocols, all of which inherit the non-negative geometry. Rectified parallel cameras place ZDP at infinity, so every finite depth yields $d=fB/z \ge 0$ by construction, and nothing within the standard pipeline can violate, or even measure, a negative disparity. To measure it, we propose ZDPShift, a benchmark of $21{,}495$ stereo pairs from seven cinematographer-authored open movies, each frame rendered at five zero-disparity-plane positions with dense signed ground truth. Six state-of-the-art image and video stereo matching models collapse once the plane moves. On identical scene content, FoundationStereo goes from $2.24$ px EPE to $75.33$ px, with every backbone leaving roughly half of all pixels exceeding a three-pixel disparity error. What is missing, however, is not the underlying matching capability. % The capability itself, however, is already present. Training on supervision synthesized from SceneFlow, which adds no new data or parameters, keeps the error flat across the signed range. Training only the decoder, with the pretrained matching features frozen, performs comparably across all six backbones, with EPE jittering within $0.2$ px. Thus, the pretrained features already extend to the negative regime they were never trained on, and only the output convention discarded it. Meanwhile, positive-regime accuracy on KITTI, Middlebury, ETH3D, and Sintel is largely preserved.

Figures & tables

Panel of Figure 1 — Figure 1: Moving the zero-disparity plane (ZDP) breaks existing stereo matching models. One scene rendered at five ZDP shifts $\Delta$ . The disparity shifts from positive only (red, in front of the ZDP) to predominantly negative (blue, behind it). We show that SOTA FoundationStereo (the 3 rd row) cannot express the negative regime. Our method (the 4 th row) restores this capability.

	FoundationStereo	StereoAnyVideo
Anaglyph
Zero-shot
Ours

Figure 2: Negative disparity in released 3D films. Four stereo films (shown as anaglyph), followed by disparity predictions (red: in front of the ZDP; blue: behind). Zero-shot FoundationStereo and Stereo-Any-Video cannot represent behind-the-screen disparities. Our signed model recovers it.

Figure 3: End-point error (EPE) on varying the ZDP shift $\Delta$ . We evaluate on six image and video stereo matching models and our signed RAFT-Stereo (solid line). Every backbone degrades sharply once $\Delta$ grows, reaching $13$ – $150$ px EPE at $\Delta=+32$ , while all six of the signed versions stay within $2.9$ – $5.2$ px across the entire signed range.

Dataset	Indoor	Outdoor	Dense	Annotation	Disp. sign	# images
Sintel ( Butler et al., 2012 )	✓	✓	✓	Synthetic	pos	1,064
KITTI12 ( Geiger et al., 2012a )	✗	✓	✗	LiDAR	pos	194
Middlebury ( Scharstein et al., 2014b )	✓	✗	✗	LiDAR	pos	15
KITTI15 ( Menze and Geiger, 2015a )	✗	✓	✗	LiDAR	pos	200
SceneFlow ( Mayer et al., 2016b )	✓	✓	✓	Synthetic	pos	35,454
ETH3D ( Schops et al., 2017 )	✓	✓	✗	LiDAR	pos	27

Table 1: ZDPShift versus existing stereo datasets along the axes that matter for the negative-disparity regime: capture/render convention, ZDP coverage, disparity sign, and scale.

Panel of Figure 4 — Figure 4: One scene, five zero-disparity planes . Top $\,\to\,$ bottom: reference view, red–cyan anaglyph, prediction error. Left $\,\to\,$ right: the same scene sweeping from positive (warm color on the $3^{rd}$ row) at $\Delta=-16$ to predominantly negative (cool color on the $3^{rd}$ row) at $\Delta=+32$ .

Figure 5: ZDPShift scene gallery , left views. Representative frames span the seven sources, including Caminandes, Spring, Sprite Fright, Settlers, Agent 327, Charge, and Project Gold.

			Neg. pixels (%) at $\Delta$
Source	Scenes	Frames	$+16$	$+24$	$+32$
The Settlers	4	1,283	$54$	$59$	$63$
Sprite Fright	7	821	$26$	$33$	$38$
Spring	5	649	$34$	$39$	$44$
Agent 327	6	590	$0$	$0$	$1$
Charge	5	435	$39$	$48$	$56$

Table 2: Composition of ZDPShift (each frame rendered at five ZDP shifts) and the share of negative-disparity pixels at

\Delta>0

. Coverage varies with the content, from Agent 327’s near-screen interiors to Project Gold’s deep exteriors.

	EPE $\downarrow$ @ $\Delta$					TEPE $\downarrow$ @ $\Delta$					Bad3 $\downarrow$
Model	$-16$	$0$	$+16$	$+24$	$+32$	$-16$	$0$	$+16$	$+24$	$+32$	${+}32$
Original (zero-shot)
Video Models
DynamicStereo	$3.84$	$3.46$	$9.02$	$14.61$	$20.90$	$1.29$	$1.28$	$1.59$	$1.93$	$2.12$	$46.6\%$
BiDAStereo	$4.12$	$3.80$	$8.51$	$18.69$	$36.62$	$1.22$	$1.21$	$1.62$	$2.12$	$2.79$	$47.2\%$
StereoAnyVideo	$3.15$	$2.73$	$6.15$	$9.17$	$12.91$	$1.14$	$1.10$	$1.18$	$1.30$	$1.44$	$45.2\%$

Table 3: Quantitative comparisons on ZDPShift . In addition to Ours (full training) , Ours (frozen) keeps the pretrained matching features frozen. Zero-shot accuracy collapses as the ZDP shifts into the negative regime, while both of our signed variants stay nearly flat with unchanged temporal consistency. Best and second-best performances are annotated in red and cyan color, respectively. A light shade is applied to the cells where the negative-disparity regime collapsed significantly.

Figure 6: The negative-disparity capability surfaces almost immediately. Both axes are logarithmic, and step 0 is the pretrained model.

	Sintel				Virtual KITTI 2
Video Models	EPE $\downarrow$		TEPE $\downarrow$		EPE $\downarrow$		TEPE $\downarrow$
	Src	Ours	Src	Ours	Src	Ours	Src	Ours
DynamicStereo	$1.95$	$\mathbf{1.69}$	$1.02$	$\mathbf{0.98}$	$1.71$	$\mathbf{1.61}$	$1.54$	$\mathbf{1.52}$
BiDAStereo	$\mathbf{1.52}$	$1.72$	$\mathbf{0.92}$	$0.98$	$1.68$	$\mathbf{1.64}$	$\mathbf{1.49}$	$1.52$
StereoAnyVideo	$\mathbf{1.47}$	$1.53$	$\mathbf{0.86}$	$\mathbf{0.86}$	$\mathbf{1.12}$	$1.29$	$\mathbf{1.11}$	$1.22$

Table 4: Positive-regime generalization across standard benchmarks. We compare the original model (Src) against ours. Ours retains competitive positive-regime accuracy and temporal consistency, with negative capability enabled.

	Trained on all $\Delta$			$\Delta={+}32$ withheld
Model	EPE -16…+24	EPE +32	TEPE +32	EPE -16…+24	EPE +32	TEPE +32
RAFT-Stereo	4.53	4.57	1.93	4.21	4.31	1.60
IGEV-Stereo	3.64	3.73	1.48	3.89	3.96	1.41
FoundationStereo	2.50	2.84	1.03	2.64	3.05	1.04
DynamicStereo	4.28	5.02	1.14	4.02	4.95	1.11
BiDAStereo	3.92	4.38	1.15	3.90	4.36	1.14

Table 5: Generalization to an unseen shift , with

\Delta{=}{+}32

withheld from training. EPE -16…+24 denotes the mean EPE of the predictions for

\Delta\in\{-16,0,8,16,24\}

Panel of Figure 7 — Figure 7: Generation degrades in negative regimes. The band is one standard error.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Film	Clips	Median	$p_{25}$	$p_{75}$
Ocean Wonders	$5$	$100.0$	$84.7$	$100.0$
Galapagos	$4$	$100.0$	$99.9$	$100.0$
Jurassic World	$12$	$86.0$	$63.2$	$95.6$
Spider-Man	$15$	$82.5$	$66.4$	$89.6$
Avatar: The Way of Water	$7$	$78.6$	$60.0$	$86.8$
Our Winter	$1$	$73.3$	$73.3$	$73.3$

Appendix

Table A: Negative-disparity share per film , one value per clip (its median frame). Percentages are of image pixels behind the screen plane, from our corrected FoundationStereo.

Figure A: Per-frame share of pixels behind the screen in ten released stereoscopic films, ordered by median. Each dot is one frame, and boxes give the per-film quartiles; the dashed line marks half the frame. Films are authored across the whole range, and most of them place the majority of the frame in the regime that no released matcher can estimate.

	ZDP placement $t$
	$0.00$	$0.25$	$0.50$	$0.75$	$1.00$
	all in front		half-half		all behind
MEt3R $\downarrow$	$0.0808$	$0.0808$	$0.0832$	$0.0855$	$0.0909$

Appendix

Table B: Generation quality against the position of the zero-disparity plane. MEt3R (lower is better) on ten clips synthesized by StereoCrafter, sweeping the plane from entirely in front of the screen (

t=0

) to entirely behind it (

t=1

). The source frames and the depth budget are identical across columns.

Panel of Figure 16 — Figure B: Real-film demonstrations for all the backbones . Every backbone forces the scene positive on zero-shot evaluations, while our version recovers the signed structure.

	EPE $\downarrow$ @ $\Delta$					Bad3 $\downarrow$ @ ${+}32$
Model	$-16$	$0$	$+16$	$+24$	$+32$
VGGT (oracle-aligned)	$17.66$	$17.38$	$17.57$	$17.66$	$17.52$	$64.2\%$
MASt3R (matched px)	$11.69$	$11.53$	$11.50$	$12.15$	$12.06$	$22.6\%$

Appendix

Table C: Zero-shot non-stereo geometry models on the ZDPShift test split (EPE per

\Delta

in native px; bad-3 at

\Delta=+32

). VGGT uses the per-frame oracle affine alignment; MASt3R is evaluated on its matched pixels. Both are flat across

\Delta

but far from the sub-pixel regime of Table 3 .

	Training		EPE $\downarrow$ @ $\Delta$
Trainable set	Params	%	$-16$	$0$	$+16$	$+24$	$+32$	Mean
FoundationStereo
none (zero-shot)	$0$	$0$	$2.98$	$2.24$	$13.06$	$58.67$	$75.33$	$30.46$
final-conv bias	$1$	${<}0.01$	$4.33$	$2.96$	$6.32$	$9.18$	$12.50$	$7.06$
final conv	$1{,}153$	$0.003$	$4.23$	$2.93$	$6.17$	$8.61$	$10.42$	$6.47$
disparity head	$0.43$ M	$1.13$	$4.17$	$2.65$	$4.05$	$4.66$	$5.15$	$4.14$

Appendix

Table D: How much of the network has to adapt. Each row fine-tunes the same recipe with a different set of parameters left trainable, from a single bias up to the whole network. EPE in native pixels under the protocol of Table 3 .

		EPE $\downarrow$ @ $\Delta$
Backbone	Condition	$-16$	$0$	$+16$	$+24$	$+32$
FoundationStereo	zero-shot	$\mathbf{2.91}$	$2.15$	$13.03$	$58.80$	$75.53$
	+ re-position, fixed $2.9\%$	$4.35$	$2.60$	$3.94$	$4.48$	$4.35$
	Ours (full tuning)	$2.97$	$\mathbf{1.83}$	$\mathbf{2.67}$	$\mathbf{2.78}$	$\mathbf{2.90}$
StereoAnyVideo	zero-shot	$\mathbf{3.15}$	$\mathbf{2.73}$	$6.15$	$9.17$	$12.91$
	+ re-position, fixed $2.9\%$	$3.70$	$3.06$	$3.58$	$3.79$	$3.93$

Appendix

Table E: Re-positioning the plane by hand, as an alternative to signed estimation. EPE in native pixels on the full ZDPShift benchmark. Fixed

2.9\%

shifts by a constant fraction of image width, the comfortable-viewing budget, which is what a practitioner can apply without knowing the offset.

	EPE $\downarrow$ @ $\Delta$					Bad3 $\downarrow$ @ $\Delta$
Model	$-16$	$0$	$+16$	$+24$	$+32$	${+}32$
IGEV (data-only)	$1.03$	$1.16$	$2.42$	$3.33$	$4.89$	$8.0\%$
FoundationStereo (data-only)	$0.70$	$0.80$	$5.09$	$7.87$	$11.11$	$41.1\%$

Appendix

Table F: Data-only ablation of the one-sided cost-volume backbones , i.e. the ZDPShift data recipe without the signed cost volume. EPE (px) per

\Delta

and bad-3 at

\Delta=+32

. Compare against the signed-volume rows of Table 3 .

Variant	$\Delta{=}{-}16$	$\Delta{=}0$	$\Delta{=}{+}16$	$\Delta{=}{+}24$	$\Delta{=}{+}32$	$\overline{\text{EPE}}$
Full recipe (multiple $\Delta$ + SceneFlow)	$\mathbf{0.92}$	$\mathbf{0.95}$	$\mathbf{0.97}$	$\mathbf{0.99}$	$\mathbf{1.01}$	$\mathbf{0.97}$
$-$ R1 ( $\Delta=0$ only)	1.31	0.95	6.67	12.18	15.61	7.35
$-$ R2 (no SceneFlow rehearsal)	$1.36$	$1.65$	$1.60$	$1.65$	$1.71$	$1.59$

Appendix

Table G: Leave-one-out ablation of the recipe ingredients on the in-domain Llamigos test split (same

3

scenes /

95

frames per

\Delta

as Table 3 ). Each row removes one ingredient and reports mean EPE per

\Delta

, and the

\Delta

-mean.

		RAFT-Stereo			IGEV-Stereo			FoundationStereo
Source / scene	#fr	ZS	Full	Frz	ZS	Full	Frz	ZS	Full	Frz
Agent 327	590	2.42	3.93	3.38	2.25	2.43	2.38	1.61	1.91	1.87
A327_02_01_A-car_enter	94	1.15	4.96	3.48	1.16	1.18	0.70	0.47	0.46	0.47
A327_04_01_H-sitting	179	5.05	5.91	5.63	4.95	5.37	5.51	4.04	4.91	4.84
A327_07_04_F-wall_slam	103	0.94	1.45	1.30	0.83	0.86	0.86	0.56	0.58	0.57
A327_08_05_A-headbutt	73	2.82	6.11	4.66	2.15	2.31	2.22	1.00	1.15	1.08

Appendix

Table H: Source- and scene-level results on ZDPShift benchmark, with image matchers (

\Delta=+32

		DynamicStereo			BiDAStereo			StereoAnyVideo
Source / scene	#fr	ZS	Full	Frz	ZS	Full	Frz	ZS	Full	Frz
Agent 327	590	1.63	2.10	1.69	1.68	2.18	1.92	1.39	1.57	1.54
A327_02_01_A-car_enter	94	0.93	2.25	1.23	1.12	1.28	1.00	0.67	0.74	0.73
A327_04_01_H-sitting	179	4.25	4.21	4.10	4.22	4.48	4.56	4.12	4.40	4.35
A327_07_04_F-wall_slam	103	0.54	0.68	0.53	0.58	0.73	0.61	0.46	0.57	0.50
A327_08_05_A-headbutt	73	2.44	3.40	2.53	2.61	4.81	3.75	1.59	2.09	2.09

Appendix

Table I: Source- and scene-level results on ZDPShift benchmark, with video matchers (

\Delta=+32

Stereo Matching

Stereo matching estimates depth by finding corresponding points in paired images and measuring their disparity.

See all 56 papers

Jun 30, 2026

LiteMatch: Lightweight Zero-Shot Stereo Matching via Cost Volume Stabilization

Despite rapid progress in learning-based stereo matching, high accuracy is often achieved at the cost of heavy backbones and computationally intensive 3D cost volume processing, resulting in substantial memory and runtime overhead. More critically, these methods frequently struggle to generalize across domains, limiting their practical deployment. We present LiteMatch, a lightweight stereo matching framework that achieves strong zero-shot generalization through cost volume stabilization-without expensive 3D convolutions. LiteMatch employs two complementary encoders: a Cross-View Correspondence Encoder (CVCE) to capture global cross-view interactions, and a High-Frequency Encoder (HFE) that enhances fine structural details via FFT-based frequency cues. To stabilize the cost volume, we introduce the Cost Volume Consistency Loss (CVC-Loss), a voxel-wise binary cross-entropy objective applied to softmax-normalized cost distributions. By encouraging sharp and unimodal disparity probabilities, CVC-Loss promotes stable cost distributions and enables rapid convergence. A lightweight refinement module further produces sharp full-resolution disparities with low-iteration updates, avoiding heavy recurrent refinement. With a flexible design ranging from 3.36M to 9.58M parameters, LiteMatch achieves exceptional zero-shot generalization, delivering competitive EPE and D1 performance across Scene Flow, KITTI, Middlebury, ETH3D, and DrivingStereo. Our results establish that lightweight architectures can indeed generalize across domains without sacrificing accuracy. blue{Code}

Md Raqib Khan, Santosh Kumar Vipparthi, Subrahmanyam Murala

CVPR Lab, Trinity College Dublin, The University of Dublin, Dublin, Ireland · CVPR Lab, Indian Institute of Technology Ropar, Rupnagar, Punjab, India

Zero-Shot Learning Stereo Matching Stereo Depth Estimation

Ask PDF

Mar 2, 2026

PromptStereo: Zero-Shot Stereo Matching via Structure and Motion Prompts

Modern stereo matching methods have leveraged monocular depth foundation models to achieve superior zero-shot generalization performance. However, most existing methods primarily focus on extracting robust features for cost volume construction or disparity initialization. At the same time, the iterative refinement stage, which is also crucial for zero-shot generalization, remains underexplored. Some methods treat monocular depth priors as guidance for iteration, but conventional GRU-based architectures struggle to exploit them due to the limited representation capacity. In this paper, we propose Prompt Recurrent Unit (PRU), a novel iterative refinement module based on the decoder of monocular depth foundation models. By integrating monocular structure and stereo motion cues as prompts into the decoder, PRU enriches the latent representations of monocular depth foundation models with absolute stereo-scale information while preserving their inherent monocular depth priors. Experiments demonstrate that our PromptStereo achieves state-of-the-art zero-shot generalization performance across multiple datasets, while maintaining comparable or faster inference speed. Our findings highlight prompt-guided iterative refinement as a promising direction for zero-shot stereo matching.

Xianqi Wang, Hao Yang, Hangtian Wang +4

Huazhong University of Science and Technology · Optics Valley Laboratory

Stereo Matching Visual Prompt Learning

Ask PDF

Mar 25, 2026

WAFT-Stereo: Warping-Alone Field Transforms for Stereo Matching

We introduce WAFT-Stereo, a simple and effective warping-based method for stereo matching. WAFT-Stereo demonstrates that cost volumes, a common design used in many leading methods, are not necessary for strong performance and can be replaced by warping with improved efficiency. WAFT-Stereo ranks first on ETH3D (BP-0.5), Middlebury (RMSE), and KITTI (all metrics), reducing the zero-shot error by 81% on ETH3D, while being 1.8-6.7x faster than competitive methods. Code and model weights are available at https://github.com/princeton-vl/WAFT-Stereo.

Yihan Wang, Jia Deng

Department of Computer Science, Princeton University

Stereo Matching Stereo Vision Stereo Depth Estimation

Ask PDF

Abstract

Stereo Matching

Explore similar work

LiteMatch: Lightweight Zero-Shot Stereo Matching via Cost Volume Stabilization

PromptStereo: Zero-Shot Stereo Matching via Structure and Motion Prompts

WAFT-Stereo: Warping-Alone Field Transforms for Stereo Matching