BINO: Encoder Centric Self Supervised Stereo With Native Pair Input

Haokun Zhou

cs.CVMar 29, 2026

BINO: Encoder Centric Self Supervised Stereo With Native Pair Input

Authors: Haokun Zhou

Organizations: Imperial College London

Abstract

Stereo needs features that preserve fine cross view correspondence rather than only semantic similarity. Recent self supervised vision models transfer well, but they are not built for this goal, and geometry focused methods often rely on a binocular decoder or another explicit linkage module during pretraining. BINO asks whether strong binocular structure can instead be learned inside a compact encoder. It does this by fusing the rectified pair at the input stage, forming stereo micro cell tokens, and using a row aware patch phase positional encoding. Training uses one view masked token only distillation together with occlusion and view specific appearance mismatch. In a strict low resource setting with pretraining only on KITTI object, BINO gives the best frozen descriptor results under a no linkage probe among all compared baselines on proxy dense stereo, hard negative retrieval, and KITTI Stereo~~2012 disparity. With the same lightweight stereo head for every encoder, it stays near CroCo~~v2 while using a much smaller encoder. Supplementary transfer experiments on KITTI Stereo~2015 show the same qualitative trend. These results suggest that much of the cross view reasoning often assigned to a separate linkage module can be learned inside a compact and reusable encoder.

Figures & tables

Variant	train_s	params (M)	Ntok	PCK@0 (%)	PCK@1 (%)	PCK@2 (%)	EPE
Interleave, stride 1, width 4	162.3	2.233	128	75.96	87.29	91.76	0.557
Concat, width 4	109.9	2.233	128	71.19	84.64	89.92	0.670
Interleave, width 2	289.6	2.229	256	74.96	86.59	91.03	0.591
Interleave, width 8	95.3	2.242	64	73.28	86.01	90.89	0.612
Interleave, width 1	350.3	2.226	512	60.24	77.28	85.08	0.959
Interleave, stride 2, width 4	160.0	2.233	128	74.00	86.48	91.39	0.587

Table 1: Architecture ablations on the controlled dual-view benchmark. Interleave fusion is clearly preferable to left-right concatenation. The best tokenization is the proposed 4

\times

2 stereo micro-cell per view, implemented here as width 4 on the fused image.

Variant	params (M)	train_s	PCK@0 (%)	PCK@1 (%)	PCK@2 (%)	EPE
1D positional embedding	2.233	161.0	76.07	87.37	91.80	0.553
Factorized 2D row+col	2.149	159.6	75.24	87.03	91.40	0.578
Full 2D grid	2.233	159.4	75.61	87.17	91.76	0.559
De-interleaved center 2D	2.149	159.4	72.97	84.51	89.51	0.674
Patch-phase 2D (ours)	2.139	159.8	75.90	88.76	92.95	0.507

Table 2: Positional encoding ablations under the selected interleave 4

\times

2 architecture. The proposed patch-phase 2D encoding yields the best error profile while using fewer parameters than the 1D baseline.

Variant	train_s	params (M)	PCK@0 (%)	PCK@1 (%)	PCK@2 (%)	EPE
Token only, uniform, EMA, scheduled mask	188.5	2.139	79.46	90.85	93.68	0.453
Token only, weighted, EMA, scheduled mask	187.9	2.139	79.11	90.18	93.15	0.479
Token+CLS, uniform, EMA, scheduled mask	188.7	2.139	76.72	88.87	92.48	0.528
Token+CLS, weighted, EMA, scheduled mask	191.6	2.139	75.51	87.83	91.84	0.565
Token+CLS, weighted, EMA, fixed mask 0.50	173.0	2.139	74.64	87.17	91.50	0.588
Token+CLS, weighted, EMA, mask both views	192.6	2.139	71.07	88.33	92.04	0.600

Table 3: Objective ablations under the selected architecture. Token-level uniform distillation is best on the primary dual-view spatial benchmark. Global CLS supervision weakens dense matching, and masking both views destroys the intended cross-view completion behavior.

Variant	train_s	params (M)	HARD_S1 PCK@1 (%)	HARD_S1 EPE	HARD_S2 PCK@1 (%)	HARD_S2 EPE	EASY_S1 PCK@1 (%)	EASY_S1 EPE
Easy only	176.9	1.670	89.63	0.518	86.01	0.613	95.56	0.293
Occlusion only	175.1	1.670	90.35	0.476	87.83	0.530	95.75	0.268
Photometric+noise only, shared	180.9	1.670	90.43	0.486	87.41	0.554	95.75	0.286
Occlusion+photometric, shared	179.3	1.670	90.65	0.462	88.29	0.510	95.83	0.261
Full hard, shared perturbation	181.9	1.670	90.79	0.456	88.27	0.509	95.85	0.261
Full hard, independent perturbation	182.0	1.670	91.17	0.448	88.30	0.513	96.43	0.247

Table 4: Data and robustness ablations. HARD_S1 is the main target benchmark. HARD_S2 increases the displacement range and acts as an out-of-distribution test. EASY_S1 is a clean reference setting.

Method	PCK@1tok (%)	EPE (px)
OURS	84.07	11.82
CroCo-v1 (sincos)	52.36	18.68
CroCo-v2 (RoPE)	79.09	16.53
iBOT-style	60.15	29.49
DINOv2-style	60.15	29.43
DINOv3-style	71.55	20.11

Table 5: Proxy dense stereo on KITTI object using StereoSGBM reference disparities. This experiment provides a large real image probe of descriptor quality without requiring stereo ground truth. Higher PCK@1tok and lower EPE are better.

Method	Top-1 (%)	Top-5 (%)	Hard@1 (%)	Hard@5 (%)	Margin
OURS	65.2	82.4	76.1	88.4	+0.0007
CroCo-v1 (sincos)	45.1	69.1	59.1	79.5	-0.0015
CroCo-v2 (RoPE)	56.3	78.6	69.2	85.7	-0.0011
iBOT-style	12.4	32.9	24.1	51.1	-0.0009
DINOv2-style	14.5	32.2	25.1	49.2	-0.0040
DINOv3-style	37.9	62.1	51.7	74.0	-0.0035

Table 6: Hard-negative retrieval on KITTI object. Retrieval is evaluated on 1,500 left-right pairs. Positives are true left-right matches. Negatives are mined from a hard random subset. Top-1 and Hard@1 are the most informative metrics. The mean margin is computed as positive similarity minus the strongest mined negative.

Method	GT WTA EPE	GT SGMLOC EPE	GT SGMLOC Bad@1tok (%)	GT+LR SGMLOC EPE	GT+LR SGMLOC Bad@1tok (%)	LRkeep (%)
OURS	9.74	4.62	2.19	4.57	1.91	92.5
CroCo-v1 (sincos)	14.76	10.25	21.36	10.22	21.29	96.8
CroCo-v2 (RoPE)	14.32	5.93	8.37	5.70	7.85	87.0
iBOT-style	28.78	7.63	11.93	7.55	11.57	71.7
DINOv2-style	34.07	8.44	15.68	8.39	15.44	67.1
DINOv3-style	19.53	8.01	14.22	7.98	14.03	80.9

Table 7: Real ground-truth disparity on KITTI Stereo 2012 validation under the strict frozen no-linkage descriptor probe. GT-only uses all ground-truth-valid tokens. GT+LR uses only tokens that also pass the left-right consistency check. SGMLOC denotes four-direction SGM on the encoder cost volume followed by local soft refinement. Lower EPE and Bad@1tok are better. Higher LRkeep is better.

Method	Encoder params (M)	GT WTA EPE	GT WTA Bad@1tok (%)	GT SGMLOC EPE	GT SGMLOC Bad@1tok (%)	GT+LR SGMLOC EPE	GT+LR SGMLOC Bad@1tok (%)	LRkeep (%)
OURS	1.339	4.423	0.78	4.258	0.70	4.215	0.56	99.4
CroCo-v1 (sincos)	3.337	5.095	2.99	4.759	2.65	4.422	1.64	77.6
CroCo-v2 (RoPE)	3.337	4.300	0.46	4.285	0.38	4.291	0.38	99.7
iBOT-style	2.020	5.606	4.04	5.742	3.64	5.718	3.62	97.7
DINOv2-style	2.022	5.593	4.02	6.232	3.54	6.210	3.49	99.3
DINOv3-style	1.928	5.684	4.63	5.806	3.65	5.803	3.65	99.4

Table 8: Shared-head benchmark on KITTI Stereo 2012. Each encoder is frozen after self-supervised pretraining on KITTI object. The same 0.086M-parameter stereo head is trained on KITTI Stereo 2012 training ground truth and evaluated on the validation split. Lower EPE and Bad@1tok are better. Since the head is identical for all encoders, differences are attributable to the encoder representations rather than the matching head.

Method	Large-disparity LOCAL Bad@1tok (%)
Ours (native fused)	15.53
Ours (export control)	21.77
CroCo-v1 (sincos)	16.29
CroCo-v2 (RoPE)	17.66
iBOT-style	20.93
DINOv2-style	21.23

Table 9: Large-disparity native pair-input stress test on KITTI Stereo 2012 validation. The subset is defined as the top quartile of token disparities on the evaluation grid. All methods use the same 0.687M pair-conditioned fused-grid decoder. Ours is evaluated in its native pair-input mode on the actual stereo pair

(L,R)

, while the export control uses the duplicated-input descriptor-export pathway based on

(I,I)

. Lower is better.

Method	Clean EPE	Clean Bad@1tok (%)	Occ40 EPE	Occ40 Bad@1tok (%)	OccAUC EPE	OccAUC Bad@1tok (%)	$\Delta$ 40 EPE	$\Delta$ 40 Bad@1tok (%)
Ours	5.403	4.37	6.886	9.74	6.284	7.50	1.483	5.37
CroCo-v1 (sincos)	5.360	4.70	7.081	9.88	6.366	7.89	1.721	5.18
CroCo-v2 (RoPE)	5.028	3.53	8.741	19.74	7.380	13.71	3.713	16.22
iBOT-style	5.831	5.06	10.725	25.71	9.032	18.39	4.894	20.65
DINOv2-style	5.855	5.21	8.625	17.97	7.619	13.46	2.770	12.76
DINOv3-style	5.562	4.58	8.928	18.69	8.218	16.25	3.365	14.11

Table 10: Native pair-input local disparity readability under unilateral one-view occlusion. The same weak local 2D head (0.242M parameters) is trained on clean KITTI Stereo 2012 training pairs for all methods. At test time, one view is corrupted by token-aligned block occlusion with area ratio from 10% to 40%. For Ours , the actual stereo pair

(L,R)

is passed through the binocular encoder and the fused token sequence is de-interleaved into its two phase maps. Lower is better. OccAUC is the average across the nonzero corruption severities, and

\Delta

40 is the degradation from clean to 40% occlusion.

Head	Params (M)	Top-1 (%)	Top-3 (%)	MRR	LOCAL EPE	LOCAL Bad@1tok (%)	BND EPE	BND Bad@1tok (%)	LARGE EPE	LARGE Bad@1tok (%)
Linear	0.032	69.1	97.6	0.830	6.381	6.23	6.676	8.46	9.668	22.30
Tiny64 $\times$ 1 $\times$ 1	0.106	78.6	97.9	0.881	5.745	5.67	6.218	8.14	8.953	21.39
Small96 $\times$ 1 $\times$ 2	0.309	77.3	97.8	0.873	5.689	5.52	6.094	7.90	8.670	21.08
Medium128 $\times$ 2 $\times$ 2	0.687	78.2	98.9	0.880	5.591	5.04	5.972	7.22	8.236	19.41
Large192 $\times$ 3 $\times$ 3	2.186	80.0	99.0	0.891	5.604	5.44	5.990	7.75	8.598	20.93
XLarge256 $\times$ 4 $\times$ 4	5.045	80.4	98.9	0.893	5.428	5.03	5.813	7.22	8.472	19.71

Table 11: Native pair-input Ours under increasing decoder capacity. The encoder is frozen and evaluated on the actual stereo pair

(L,R)

. Lower EPE and Bad@1tok are better. Higher Top-1, Top-3, and MRR are better.

Variant	Total params (M)	GT WTA EPE	GT WTA Bad@1tok (%)	GT SGMLOC EPE	GT SGMLOC Bad@1tok (%)	GT+LR SGMLOC EPE	GT+LR SGMLOC Bad@1tok (%)	Top-1 (%)	Hard@1 (%)	Margin
Base model	1.771	5.184	14.66	2.202	1.73	2.192	1.64	64.3	76.3	0.0026
+ register tokens	1.772	5.134	14.44	2.266	1.88	2.256	1.78	58.9	72.8	0.0015
+ SwiGLU FFN	1.779	5.003	13.83	2.144	1.43	2.131	1.33	64.2	76.9	0.0028
+ dual token/global heads	1.797	4.450	12.38	2.308	2.01	2.300	1.96	70.8	82.8	0.0034

Table 12: Parameter-controlled refinements around the real-data recipe. All variants keep approximately the same total parameter count by compensating the feed-forward ratio. Lower stereo error and higher retrieval scores are better.

Figure 1: Emergence of epipolar structure across depth in the fused DePos representation. Same-row concentration and ground-truth window mass increase steadily from the embedded input state to the final DePos readout on both controlled synthetic stereo and KITTI Stereo 2012 validation. Dashed lines indicate chance levels. The final DePos readout is marked explicitly to show that applying DePos preserves the already-emergent geometry rather than altering it substantially.

	Synthetic stereo				KITTI Stereo 2012 val
Layer	RowConc	GT@1	MRR	Entropy	RowConc	GT@1	MRR	Entropy
layer0_raw	0.1036	0.0990	0.2200	3.4568	0.0956	0.0718	0.3875	3.5953
layer1_raw	0.1408	0.1563	0.2642	2.9820	0.1402	0.1416	0.4406	3.1353
layer2_raw	0.1612	0.1753	0.2760	2.8855	0.1677	0.1691	0.4317	2.9826
layer3_raw	0.1602	0.1782	0.2823	2.9319	0.1701	0.1726	0.4361	2.9892
layer4_raw	0.1647	0.1806	0.2852	2.9476	0.1784	0.1778	0.4285	2.9695

Table 13: Layerwise emergence of epipolar structure in the original fused DePos representation. Metrics are computed from frozen token similarities between even-phase query tokens and odd-phase candidate tokens. Higher RowConc , GT@1 , and MRR are better. Lower Entropy is better. For brevity, the table reports the final DePos readout rather than both final raw and final DePos; the two are numerically almost identical, as discussed in the text.

Figure 2: Counterfactual evidence for binocular causal structure. Left: retention relative to the original DePos readout under replace-right and row-shuffle-right . Row shuffling preserves same-row concentration but sharply reduces ground-truth-window retention, showing that the learned representation depends on the ordered horizontal structure of the complementary view rather than on a generic row prior. Right: under the duplicate-left counterfactual, zero-disparity metrics rise to several times chance, indicating that the representation shifts coherently toward zero disparity when the two views are made identical.

	Synthetic stereo				KITTI Stereo 2012 val
Condition	RowConc	Tgt@1	MRR	Entropy	RowConc	Tgt@1	MRR	Entropy
Original (raw)	0.1726	0.1782	0.2835	2.9708	0.1787	0.1781	0.4284	2.9672
Original (DePos)	0.1722	0.1783	0.2835	2.9695	0.1782	0.1784	0.4283	2.9658
Replace-right	0.1530	0.1603	0.2754	3.0595	0.1550	0.1612	0.4281	3.0712
Row-shuffle-right	0.1723	0.1163	0.2339	3.2364	0.1762	0.1047	0.2607	3.2632
Duplicate-left (zero)	0.1676	0.2603	0.6638	2.8850	0.1745	0.2650	0.7340	2.9336

Table 14: Final-layer counterfactual analysis of the original fused DePos representation. The original condition uses the true stereo pair and ground-truth disparity target. Duplicate-left uses a zero-disparity target because both inputs are identical. Higher RowConc , Tgt@1 , and MRR are better. Lower Entropy is better.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Top-1 (%)	Top-5 (%)	GT WTA EPE	GT SGM EPE
OURS	70.0	100.0	12.728	6.728
CroCo-v1 (sincos)	35.0	70.0	9.076	10.849
CroCo-v2 (RoPE)	68.3	98.3	15.400	7.605
iBOT-style	18.3	41.7	28.327	8.151
DINOv2-style	16.7	55.0	23.254	9.985
DINOv3-style	55.0	90.0	15.059	8.437

Appendix

Table 15: Supplementary transfer check on KITTI Stereo 2015 using a deterministic held-out split of the public training scenes. All encoders use the same low-resource pretraining recipe as in the main KITTI Stereo 2012 study.

Method	Encoder params (M)	GT WTA EPE	GT WTA Bad@1tok (%)	GT SGMLOC EPE	GT SGMLOC Bad@1tok (%)	GT+LR SGMLOC EPE	GT+LR SGMLOC Bad@1tok (%)	LRkeep (%)
OURS	1.339	4.742	0.95	3.888	1.29	3.831	1.06	99.7
CroCo-v1 (sincos)	3.337	5.022	1.97	4.313	2.05	4.209	1.63	98.4
CroCo-v2 (RoPE)	3.337	4.637	0.71	4.224	0.92	4.190	0.77	99.7
iBOT-style	2.020	5.316	2.58	4.599	2.58	4.598	2.57	98.9
DINOv2-style	2.022	5.719	4.21	5.377	4.13	5.377	4.13	100.0
DINOv3-style	1.928	5.294	2.47	6.401	3.63	6.401	3.63	100.0

Appendix

Table 16: Supplementary shared-head transfer check on KITTI Stereo 2015 using a deterministic held-out split of the public training scenes. Each encoder is frozen after the same low-resource pretraining recipe as in the main KITTI Stereo 2012 study, and the same lightweight stereo head is trained for all methods. Lower EPE and Bad@1tok are better.

Visual Representation Learning

Visual representation learning develops image and video features that capture useful structure for recognition, retrieval, spatial reasoning, and action.

See all 224 papers

NBS: No Bias Stereo
Aug 28, 2026Vage Taamazyan, Zhuowen Shen, Stefan Hinterstoisser +6Stereo Matching Stereo Depth Estimation
Intrinsic (Google) · Texas A&M University
MAGIC: Learning from Visibility Asymmetry for Unsupervised Stereo Matching
Aug 14, 2025Peng Xu, Zhiyu XiangStereo Matching Occlusion Reasoning
College of Information Science and Electronic Engineering (ISEE) Zhejiang University, China
Lite Any Stereo V2: Faster and Stronger Efficient Zero-Shot Stereo Matching
Jun 23, 2026Junpeng Jing, Ronglai Zuo, Zhelun Shen +5Stereo Matching Stereo Depth Estimation
Imperial College London, London SW7 2AZ, United Kingdom

Aug 28, 2026J/K move · Enter open · S save

NBS: No Bias Stereo

Vage Taamazyan, Zhuowen Shen, Stefan Hinterstoisser, Alberto Dall'Olio+5

Intrinsic (Google) · Texas A&M University

Stereo Matching Stereo Depth Estimation 3D Reconstruction

Open paper AskPDF arXiv

Stereo reconstruction is one of the last remaining Computer Vision tasks where all state-of-the-art methods employ a heavy architectural inductive bias. Even though it has been demonstrated that the task can be solved using general-purpose methods, it is widely believed that inductive biases in stereo are strictly necessary for both high-quality results and computational efficiency. We challenge this paradigm. In this paper, we demonstrate that both state-of-the-art accuracy and superior runtime efficiency are achievable with a model completely devoid of architectural inductive biases, relying instead on a simple, end-to-end Vision Transformer. By training on massive synthetic datasets, we show that pure data-driven learning can surpass explicitly engineered geometry. This work proves that explicit inductive biases are no longer a prerequisite for stereo matching, ultimately unlocking true scaling laws for continuous improvement in 3D reconstruction.

Abstract

Visual Representation Learning

Explore similar work

NBS: No Bias Stereo

MAGIC: Learning from Visibility Asymmetry for Unsupervised Stereo Matching

Lite Any Stereo V2: Faster and Stronger Efficient Zero-Shot Stereo Matching