Which depths in a new camera view are determined by finitely many source-depth measurements? Pointwise uncertainty bounds do not answer this question: unobserved geometry between samples can change which surface is encountered first. We study a bounded Lipschitz class of continuous source inverse-depth fields with interval observations and known camera pose. Building on classical Lipschitz envelopes, we characterize the exact set of attainable first-hit depths and construct a globally admissible scene for every value in that set. The set can be disconnected. We also give a necessary and sufficient test for several proposed depths to be first hits of one shared scene. An epipolar reduction uses scalar cone inequalities without enumerating surfaces. In a rectified model with source spacing h, interval half-width ε, baseline b, and slope bound L, we prove a sharp worst-case inverse-depth diameter of max{Lh+2ε,2ε/(1−bL)} when bL<1, subject to sufficient depth and domain range for attainment. Exact samples still give the sharp bound Lh at bL=1, while above this threshold ambiguity need not vanish as sampling becomes arbitrarily dense. Any positive measurement uncertainty can destroy uniform stability at the critical threshold. We provide constructive examples, an exact rational validation oracle, and a frozen ten-scene evaluation using reference, stereo, and learned-depth observations. The experiments separate solver correctness under the scene assumptions from failure of those assumptions: fixed learned-depth bands often yield narrow but incorrect sets. The results identify what a specified scene class permits; they do not infer a continuum regularity or physical-scene guarantee from sampled predictions.
Figures & tables
Figure 1: Identifying the first visible surface from finite depth samples. (a) Problem: exact depth one at every integer source pixel (dots) is matched by a smooth surface (solid) whose first hit along a target ray is at depth 1.659 , while a discrete interval splat returns [1,1] (dashed); pointwise intervals do not constrain geometry between samples. (b) Mechanism: with a slope bound L , the lower and upper Lipschitz envelopes f,f along the ray’s epipolar trajectory give the exact set of attainable first hits, Q={q≥q−:q≤f(x(q))} (black bars; example of Eq. 10 ). Unlike pointwise bounds or an interval hull, Q excludes depths that are necessarily occluded and impossible interior bands (hatched); each q∈Q is realized by an explicit witness scene f(q) (dotted cone). (c) Theory (schematic, not data): on a rectified scanline with spacing h and noise half-width ε , the sharp worst-case diameter is Lh for exact samples up to the fold threshold κ=1 and max{Lh+2ε,2ε/(1−κ)} with noise; above the threshold, denser sampling cannot remove ambiguity.
Figure 2: Exact support and shared visibility. Left: two exact observations leave a disconnected first-hit support; an interval hull includes an impossible band. Right: individually possible, mutually Lipschitz-compatible depths can still be impossible as joint first hits. The forced lower cone creates an earlier occluder. These are analytic constructions, not measured prevalence.
Figure 3: Resolution, noise, and folds. Analytic extremizers and computed identified sets show the two sharp subcritical regimes, critical sensitivity to positive noise, and persistent ambiguity above the fold threshold. Markers are computed support diameters for the proved analytic families, with lines connecting them; dashed limits are labeled. Domain/range clipping is avoided in sharpness cases.
Source observations
Queries
In class (%)
Accept at 5% (%)
Digital in-class
Measured > 5% errors
Reference bands
14,700
22.97
30.78 [19.44, 44.33]
4,170/4,170
0/5,380
StereoSGBM
4,900
4.05
33.82 [21.29, 46.69]
265/265
95/1,986
Depth Anything V2
4,900
3.87
2.82 [0.80, 5.86]
225/225
169/169
ZoeDepth
4,900
4.61
9.92 [5.12, 15.77]
295/295
409/503
Table 1: Completed ten-scene primary evaluation. In-class and acceptance rates are equal-scene means; brackets are descriptive 95% scene-bootstrap intervals. Coverage and error columns are pooled counts with explicit denominators. Measured error uses accepted queries with an available right-view reference. Reference bands (E1a) have three noise replicates; the other studies have one. Digital membership is an oracle evaluation label, not a selection rule. Ten SGBM in-class records have no supplied observations. All primary cells have Lβ=1 , hence connected support. These empirical rates do not establish a physical-scene certificate.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Study
Queries
Applicable
In-class contained
Accept at 5% [95% interval]
Accepted > 5% errors
O1 primary cell
3,000
98.40%
2,382/2,382
98.40 [96.20, 100.00]
0/2,952
O1 sensitivity settings
807,000
96.20%
562,512/562,512
80.92 [75.73, 85.78]
0/653,035
O2 correct pose
2,460
0.00%
–
0.00
–
O2 20 perturbations
49,200
0.00%
–
0.00
–
O2-w correct pose
2,700
100.00%
–
100.00
0/2,700
O2-w rotation 0.1° (4)
10,800
100.00%
–
100.00
0/10,800
Appendix
Table H.1: Optional studies on the held-out data ( Appendix J ). O1 is compute-bounded to four complete windows per scene; its truth is a bilinear digital scene, so containment and errors are conditional digital checks, not physical-scene claims. Acceptance is an equal-scene mean with a descriptive 95% scene-bootstrap interval; containment and error columns are pooled counts. O2 uses measured target truth under an intentionally misspecified pose; every O2 query was refused, so it supports no sensitivity claim. O2-w is a labelled post-hoc wide-window variant ( Appendix J ); its rows pool the listed number of conditions over the nine evaluable scenes, give the equal-scene acceptance mean without an interval, and count accepted queries with measured error above 5%. No nonprimary setting was dropped.
Check
Cases/checks
Failures
Largest discrepancy or scope
Independent rational support oracle
600
0
3.464×10−14 endpoint error
Generated one-dimensional fields
500
0
Reference first-hit inclusion
Generated two-dimensional fields
300
0
Reference first-hit inclusion
Constructed global witnesses
2,064
0
1.128×10−13 constraint residual
Subcritical sharpness
72
0
8.794×10−14 diameter error
Critical exact-sample sharpness
30
0
7.319×10−14 diameter error
Appendix
Table I.1 : Completed numerical verification. Rows share an implementation and some analytic constructions; they are not independent statistical replications. The fast implementation also passes the same 600 rational cases, which are counted only once in the table. Separate infeasible-input and domain checks exercise refusal paths.
Primary development cell
Planned
Feasible
Complete ref.
In class
Misses/in class
Mean R∗
Adirondack
700
265
335
165
0/165
0.862%
Motorcycle
700
270
235
160
0/160
0.943%
Pipes
700
145
165
100
0/100
1.466%
Pooled
2,100
680
735
425
0/425
1.035%
All 81 settings, pooled
170,100
60,795
59,535
31,860
0/31,860
2.091%
Appendix
Table I.2: Development results. Primary (L,h,ε,β)=(1,8,0.5,1) . R∗ is averaged only over in-class digital references. Counts of complete references and feasible classes have different denominators and neither implies the other. There are 109,305 refusals over all settings. The 170,100 records reuse three physical scenes and are not independent scenes.
Figure I.1 : Completed development diagnostic. Panel (a): feasible-query and complete-reference membership rates retain every prespecified query. Panel (b): conditional relative minimax radius, with descriptive ten-to-ninety percent quantiles across queries, not confidence intervals. Panel (c): full-information versus two-local-sample hull widths. Dense-reference membership is an oracle evaluation label, not an input to a deployment selection rule.
Source
Planned
Feasible
Digital ref.
In class
Empty band
Incompatible
No obs.
Reference bands
14,700
5,455
5,250
4,170
0
9,245
0
StereoSGBM
4,900
2,990
1,750
265
0
1,910
65
Depth Anything V2
4,900
1,880
1,750
225
2,930
90
0
ZoeDepth
4,900
3,155
1,750
295
1,560
185
0
Appendix
Table I.3: Primary accounting. Empty-band and incompatible-class counts are refusals. Missing digital references remain in the planned denominator. All four primary studies have zero numerical errors, no geometric-domain refusals, and no complete-reference range failures. Each estimator has 360 queries with a complete reference violating L ; E1a repeats these three times. Failure categories concerning the true scene can overlap, and must not be added as a partition.
Source
Sample band coverage
Digital coverage
Measured coverage
In-class digital
Reference bands
878,700/878,700
4,530/4,530
5,375/5,386
4,170/4,170
StereoSGBM
123,880/205,640
1,486/1,540
2,678/2,932
265/265
Depth Anything V2
22,705/292,900
252/1,155
299/1,839
225/225
ZoeDepth
79,955/292,900
471/1,465
1,128/3,089
295/295
Appendix
Table I.4: Containment with explicit denominators. Source counts are query-weighted supplied observation incidences with finite reference values: a source sample is repeated for its five rays. They are not independent pixels. Digital and measured target coverage use finite feasible outputs with the corresponding reference, without filtering to true class membership. The final column does condition on membership. Source coverage for empty-observation windows is undefined; vacuous constraints count only when testing mathematical class membership.
Source
Source coverage (%)
In class (%)
Accept at 5% (%)
Measured violations (%)
Reference bands
100.0 [100.0, 100.0]
23.0 [12.5, 36.0]
30.8 [19.4, 44.3]
0.0 [0.0, 0.0]
StereoSGBM
60.8 [56.4, 64.6]
4.1 [1.3, 7.7]
33.8 [21.3, 46.7]
3.7 [1.2, 6.7]
Depth Anything V2
7.3 [0.2, 17.1]
3.9 [0.0, 9.3]
2.8 [0.8, 5.9]
100.0 [100.0, 100.0]
ZoeDepth
25.7 [8.9, 44.7]
4.6 [0.0, 10.1]
9.9 [5.1, 15.8]
89.9 [76.3, 100.0]
Appendix
Table I.5: Equal-scene primary rates with descriptive 95% intervals from 10,000 paired scene resamples. The last metric excludes scenes with no accepted evaluable query: 9, 9, 8, and 10 scenes contribute, respectively. A degenerate interval at zero or one is a limitation of resampling the observed finite collection, not a statement that the population rate is known exactly. No pixel-level binomial interval is used.
Source
Common queries
Linear
PCHIP
Envelope midpoint
Minimax
Reference bands
4,530
0.0784
0.0844
0.0970
0.3119
StereoSGBM
1,087
0.4519
0.4550
0.4580
0.5837
Depth Anything V2
1,155
25.9545
25.9535
25.9516
26.0897
ZoeDepth
1,465
23.7078
23.7083
23.6663
23.6683
Appendix
Table I.6: Pooled mean relative-depth error in percent on exactly the same evaluable digital-reference queries for all four point estimators within a source observation condition. A valid feasible class and finite output from all comparators are required. Source conditions can have different common subsets and information budgets, so rows are not an accuracy leaderboard between the source estimators. Scene-wise median and 90th percentile errors and their bootstrap intervals will be released with the per-query records. The minimax estimator optimizes a worst-case completion loss; average point-accuracy superiority is not observed.
Scene
Planned
Feasible
In class
Source cover (%)
Accept at 5%
Measured > 5%
Backpack
2100
900
705
100.00
891
0/879
Bicycle1
2100
780
540
100.00
747
0/735
Cable
1500
560
450
100.00
560
0/560
Classroom1
2100
1710
1530
100.00
1704
0/1704
Couch
900
150
75
100.00
150
0/150
Flowers
1200
340
255
100.00
340
0/340
Appendix
Table I.7: Reference bands: every primary physical scene. The final column is the number of violations divided by accepted queries with a measured right-view reference. A dash denotes an undefined error rate with zero evaluable accepted outputs, not zero error.
Scene
Planned
Feasible
In class
Source cover (%)
Accept at 5%
Measured > 5%
Backpack
700
540
65
61.38
444
31/430
Bicycle1
700
395
20
46.04
321
42/310
Cable
500
290
25
57.64
190
3/189
Classroom1
700
585
125
71.72
480
0/477
Couch
300
120
0
60.16
25
0/25
Flowers
400
215
0
60.30
106
1/105
Appendix
Table I.8: StereoSGBM: every primary physical scene. The final column is the number of violations divided by accepted queries with a measured right-view reference. A dash denotes an undefined error rate with zero evaluable accepted outputs, not zero error.
Scene
Planned
Feasible
In class
Source cover (%)
Accept at 5%
Measured > 5%
Backpack
700
255
110
32.64
21
21/21
Bicycle1
700
135
0
0.02
105
103/103
Cable
500
190
115
35.13
1
1/1
Classroom1
700
650
0
0.37
8
8/8
Couch
300
5
0
0.00
0
– (0)
Flowers
400
205
0
0.22
9
9/9
Appendix
Table I.9: Depth Anything V2: every primary physical scene. The final column is the number of violations divided by accepted queries with a measured right-view reference. A dash denotes an undefined error rate with zero evaluable accepted outputs, not zero error.
Scene
Planned
Feasible
In class
Source cover (%)
Accept at 5%
Measured > 5%
Backpack
700
380
160
41.63
101
97/99
Bicycle1
700
150
0
0.00
109
109/109
Cable
500
490
70
79.29
153
62/150
Classroom1
700
630
0
0.30
2
2/2
Couch
300
35
0
0.00
17
17/17
Flowers
400
335
0
21.78
11
6/10
Appendix
Table I.10: ZoeDepth: every primary physical scene. The final column is the number of violations divided by accepted queries with a measured right-view reference. A dash denotes an undefined error rate with zero evaluable accepted outputs, not zero error.
Source
Cells
Query records
Feasible
In-class coverage
Disconnected
Numerical failures
Reference bands
81
1,190,700
514,095
321,570/321,570
113,744
0
StereoSGBM
135
661,500
436,335
95,070/95,070
65,663
0
Depth Anything V2
135
661,500
347,235
67,995/67,995
24,136
0
ZoeDepth
135
661,500
438,030
68,580/68,580
31,979
0
Appendix
Table I.11: Entire fixed sensitivity grid, including all outcomes. E1a uses three noise realizations per query and configuration; the other interfaces use one. Counts of disconnected outputs concern feasible classes, not only oracle-selected references. Gaps below the prescribed numerical tolerance are ignored for component statistics, while raw support endpoints are retained in the per-query records.
Source
Feasible
Disconnected (%)
Depth gaps (%)
Hull reduction (%)
Reference bands
8,540
92.1 [82.4, 97.6]
23.6 [20.1, 25.5]
33.7 [27.1, 39.9]
StereoSGBM
3,890
60.0 [42.9, 74.1]
13.9 [7.9, 19.0]
16.9 [8.4, 25.3]
Depth Anything V2
1,955
52.0 [26.0, 75.7]
1.0 [0.5, 1.6]
9.2 [4.0, 16.0]
ZoeDepth
3,305
47.8 [29.2, 65.5]
0.8 [0.3, 1.3]
7.8 [3.3, 13.9]
Appendix
Table I.12: Predicted folded regime, changing only L from one to four relative to each primary cell ( h=8 , β=1 , source band unchanged). Intervals are descriptive 95% scene-bootstrap intervals. Disconnectedness is a mean scene fraction of feasible outputs; gap fraction and hull reduction are means of within-scene medians. Gap fraction is excluded target-depth length divided by its sharp hull length. Hull reduction is one minus the first-hit depth-hull width divided by the all-intersection depth-hull width. No query is selected by its measured error or reference class membership. All primary cells have zero gaps and zero reduction in this comparison.
Figure I.2 : Source spacing with each primary source band, L=1 , and β=1 . Curves are equal-scene means of within-scene median relative minimax uncertainty R∗ . Shading and whiskers show descriptive 95% scene-bootstrap intervals; stars mark the frozen stride-eight primary cell. The contributing scene count can depend on feasibility and is reported in the aggregates. These are conditional uncertainty radii, not empirical calibrated risks.
Figure I.3 : Completed primary operating-point diagnostics. Left: equal-scene acceptance and measured-right error-violation rates at tolerances 2%, 5%, and 10%, with descriptive scene-bootstrap intervals. Faint points show individual scenes at 5%; empty-acceptance risks are absent. Right: a mutually exclusive partition of every planned query, averaged over ten scenes. Empty observation sets have their own category; they are not evidence of valid model bands. Green whiskers quantify nonempty verified membership and black whiskers feasible-output fraction.
Figure I.4 : Entire frozen sensitivity grid at β=1/4 . The four panels are the four source interfaces. Colors encode every fixed source band and line style the regularity bound. Each line shows the equal-scene mean of within-scene median R∗ among finite outputs as source stride changes. Undefined conditional metrics remain gaps. No setting is selected by target error. All scene counts and bootstrap intervals are in the numerical records.
Figure I.5 : Entire frozen sensitivity grid at β=1/2 , with the same panels, bands, regularity values, source spacings, and aggregation rule as Fig. I.4 . These are completion-class uncertainty radii, not calibrated physical risks. All fixed settings are retained.
Figure I.6 : Entire frozen sensitivity grid at β=1 . Black stars mark the frozen primary cells. Colors encode source bands and line styles the regularity bounds, as in Fig. I.4 . Primary-spacing intervals appear in Fig. I.2 ; every full-grid cell’s interval and contributing scene count is in the numerical records.
Model
Records
Decision mismatches
Max. rel. difference
Acceptance flips
Depth Anything V2
4,900
0
1.2×10−6
0
ZoeDepth
4,900
0
1.0×10−6
0
Appendix
Table J.1: Frozen E2 versus the fresh fp32 re-execution. Decision fields include membership, feasibility, refusal reason and component count. The relative difference covers support endpoints, radius and the reference values. Acceptance flips count changes at the 2, 5 and 10% tolerances. Every entry of Tab. 1 for these two models is therefore reproduced exactly.