acceptodds
Under review as a conference paper at ICLR 2027

Phy-DRILL: 2.5D Physical Depth Reversal Illusion against Monocular Depth Estimation

Abstract

Reliable monocular depth estimation (MDE) is crucial for the operational safety of autonomous systems, making it essential to evaluate its robustness in the physical world. Yet existing physical attacks, whether textures optimized on 2D patches or 3D meshes, alter only appearance and treat physical geometry as a fixed constraint. In this work, we reveal that parameterized geometry unlocks adversarial effects that appearance alone cannot reach. We exploit the Depth Reversal Illusion (DepthRI), a perceptual phenomenon in which protruding structures are visually misinterpreted as receding surfaces due to conflicting geometric and texture cues. We show that DepthRI deceives not only human vision but also SOTA MDE models and trace its effect to two factors: the number of vanishing points in the perspective structure, and the texture gradient that activates the illusion. Guided by these findings, we propose Phy-DRILL (Physical Depth Reversal ILLusion), which constructs 2.5D adversarial patches grounded in perspective geometric relationships and further enhances their adversarial effectiveness through targeted surface texture optimization. Our approach causes physically protruding structures to be systematically misperceived as concave regions by SOTA MDE models. Across six MDE models, the experiments show that Phy-DRILL outperforms existing attacks in both digital and physical settings and transfers from a single surrogate model to five black-box models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Physical Adversarial Attack, Monocular Depth Estimation, Depth Reversal Illusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eXjJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Phy-DRILL, a physical attack against monocular depth estimation (MDE) based on the depth-reversal illusion. The central idea is to construct a physically convex object that appears visually concave: a truncated-pyramid structure protrudes toward the camera, while perspective-consistent textures printed on its surfaces resemble a corridor receding into the distance. Phy-DRILL jointly optimizes the geometry of the truncated pyramid and its surface texture using differentiable rendering. Its objective increases the predicted depth within the object region, encourages a pronounced protruding geometry, and uses a Stable Diffusion prior to preserve a corridor-like texture. The optimized texture is printed on paper and folded into the corresponding physical structure. Experiments use Depth Anything V2 as the white-box surrogate and evaluate transfer to five additional MDE models in digitally composited KITTI scenes and controlled near-field physical environments. The reported results show stronger attack performance than the evaluated 2D and 3D adversarial-patch baselines.

Critical strengths
  1. Interesting geometry-aware attack idea. The paper presents an interesting physical attack that goes beyond conventional texture-only adversarial patches by treating physical geometry as an attack variable. Exploiting the depth-reversal illusion to make a convex object appear concave to MDE models is intuitive and provides a new perspective on the vulnerability of monocular depth cues.

  2. Physical realization and broad model evaluation. The proposed object can be fabricated using printed paper and a simple folded structure. The paper evaluates both digital and physical attacks across six MDE models and includes ablations separating the effects of geometry and texture, providing initial evidence that the attack transfers beyond the white-box surrogate.

Critical weaknesses
  1. The physical viewpoint robustness and practical deployability are not sufficiently demonstrated. The depth-reversal illusion depends on an alignment between the camera, the frustum geometry, and the perspective texture. Although the paper reports digital evaluations over a wide angular range, similar camera variations are already included in the optimization process, and the physical evaluation mainly considers frontal and ±30° views in a controlled near-field setting. It remains unclear whether the illusion persists under continuous ego-camera motion, unseen elevations and trajectories, realistic self-shadowing, or strongly oblique viewpoints where the object’s silhouette and occlusion reveal its convex geometry. Moreover, the reported relation implies a protrusion height of approximately 2 m at a 10 m observation distance, raising concerns about the practicality of the attack in realistic driving scenarios.

  2. Evaluation metric may not establish depth reversal. The current evaluation metrics do not clearly distinguish depth reversal from ordinary depth-estimation error. In the physical experiments, the ground-truth depth is approximated using the camera-to-object placement distance. However, the adversarial object is a protruding frustum whose surface pixels have different physical depths. Treating the entire object region as having a single ground-truth depth may therefore count geometrically correct predictions as errors. This issue is important because the physical setup uses a frustum height of 0.1 m at a placement distance of 0.5 m, whereas (R_a) marks pixels as affected once their error exceeds only 10% of the reference depth. In addition, the absolute-error metric (E_d) does not distinguish the intended depth overestimation from underestimation or other prediction errors. The evaluation should use dense geometry-aware ground truth and report signed depth overestimation or an explicit depth-polarity reversal rate.

  3. The evaluation is limited to frame-level depth errors and does not demonstrate a system-level safety impact. The digital experiments composite the adversarial object onto individual KITTI images, while the physical experiments evaluate isolated frames in controlled scenes. It remains unclear whether the attack remains temporally consistent as the vehicle approaches the object, or whether the induced depth error is sufficient to affect downstream obstacle detection, tracking, planning, and control. A closed-loop evaluation in CARLA or a comparable simulator, with a moving ego vehicle and an MDE-based perception and planning stack, would be necessary to substantiate the claimed autonomous-driving risk.

Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response regarding to Weakness 1.

W1. How does the method behave at longer observation distances and more oblique physical viewpoints?

Thanks for raising this concern. The larger patch still affected the depth predictions of all four evaluated models at 8.0 m, the farthest distance we tested. In the angle experiment, all evaluated models showed depth deviations at both −70° and +70°, with smaller deviations at these angles than in the frontal view.

Setup. The main physical comparison in the original submission used a 100mm-high patch observed at 0.5 m. In the additional experiment, we fabricated a larger patch from printed KT foam board and mounted it at the rear of a passenger car. The patch was 800 mm high, with base and top-face edge lengths of 722 mm and 123 mm, respectively. We captured photographs in a road scene and evaluated the MDE predictions offline. The distance experiment covered seven observation distances from 2.0 to 8.0 m. For the angle experiment, we fixed the observation distance at 4.0 m and captured one photograph every 10° from −70° to +70°. We report EdE_d, the mean absolute deviation from the paired reference prediction within the patch mask, on each model's output scale.

(1) Observation distance. All four models had smaller deviations at 8.0 m than at 2.0 m. DA-V2-Metric's EdE_d decreased steadily from 9.46 to 1.43. Metric3D showed a sharper drop from 43.66 at 2.0 m to 3.40 at 5.0 m, followed by fluctuations at longer distances and a deviation of 6.14 at 8.0 m. Table R3.1 shows how the effect changed with distance for each model.

(2) Viewing angle. The frontal view produced the largest deviation for three of the four models in Table R3.2. ZoeDepth peaked at −20°. For Metric3D, EdE_d was 5.44 at 0°, 0.58 at −70°, and 1.63 at +70°. Some intermediate views retained much of the frontal-view deviation. For example, Monodepth2's EdE_d was 7.28 at +20°, compared with 8.06 at 0°. The angle sweep covers oblique views in a road scene, and the approach sequence in W3 addresses persistence under camera motion.

(3) Patch height at longer distances. The reported 20.58% is the average height-to-distance ratio that maximized predicted depth in our simulation sweep. This is where the estimate of a roughly 2.0 m optimal height at a 10.0 m observation distance comes from. In the road experiment, the 800 mm-high patch was tested at 8.0 m, corresponding to a height-to-distance ratio of 10%. The results at that distance show that the patch can affect depth predictions at a smaller ratio than the simulated optimum.

Table R3.1. EdE_d at different observation distances.

Observation distance (m)2345678
DA-V2-Metric9.465.815.544.263.522.951.43
ZoeDepth4.242.593.270.473.270.270.08
Metric3D43.6642.925.443.405.861.636.14
Monodepth210.028.848.064.821.860.720.47

Table R3.2. EdE_d at different viewing angles at 4.0 m.

Viewing angle (°)−70−60−50−40−30−20−100+10+20+30+40+50+60+70
DA-V2-Metric1.401.231.223.193.814.104.365.544.353.183.171.420.741.291.13
ZoeDepth0.810.250.191.551.505.710.633.270.430.571.870.500.200.190.32
Metric3D0.580.470.900.961.472.862.345.444.623.414.670.990.220.801.63
Monodepth22.971.000.870.881.424.866.128.065.877.283.512.221.011.430.90
AuthorsOpenReview ↗

Response regarding to Weakness 2.

W2. How do the reference depth and evaluation metrics distinguish depth reversal from ordinary depth errors?

Thanks for raising this question. The reference for assessing depth reversal in our physical evaluation is the depth of the patch's base plane, and we chose it deliberately. The prediction must cross the base plane to count as reversal.

(1) A stricter criterion for reversal. The 2.5D patch physically protrudes toward the camera, so its protruding surface lies in front of the base plane. A depth estimation model can overestimate the surface depth and still place it in front of the base plane. The patch would then merely appear to protrude less. We count a reversal only when the predicted surface is placed behind the base plane. Moreover, once the prediction crosses the plane, the deviation measured from the base plane is smaller than the deviation measured from the true protruding surface. In this case, the base-plane reference gives a conservative measure of the signed depth overestimate.

(2) The base-plane reference is verified experimentally. To confirm that the base-plane depth is a reliable reference, we run the same depth estimation model on the original image and on a paired image in which the patch region is filled with white to remove the texture. In our validation experiment, the predicted depth of the white region agreed with the physically measured base-plane distance. Deviations (EdE_d) are computed against this paired prediction. Reversal is checked by whether the predicted patch depth goes beyond the measured base plane. We agree with the reviewer's point about RaR_a. When 0.50 m is used as the surface reference, a correct prediction of 0.40 m at a protruding point gives a deviation of 0.10 m and exceeds the 0.05 m threshold.

(3) Physical evidence. To validate that our method achieves depth reversal, we conducted an additional physical experiment that measures the spatial profile of the predicted depth across the patch, from its protruding center to its edge.

Setup: A frontal-view photograph was captured in a real road scene, with the camera 2.0 m from the patch's base plane. Within the visible patch region, we defined five concentric rectangles: one central region plus four non-overlapping bands, so that we can examine how predicted depth changes from the center to the edge. Table R3.3 reports the geometric reference depth and the mean predicted depths of three metric MDE models in each region.

Geometric reference (Real Depth): depth increases from center to edge. Because the patch physically protrudes toward the camera, its geometric reference depth increases monotonically from Region 1 (center, 1.2 m) to Region 5 (outermost band, 2.0 m), by 0.80 m overall; the four inner regions all lie in front of the 2.00 m base plane. Note that a model which merely overestimates depth while still interpreting the surface as convex would preserve this ordering — closest at the center, farthest at the edge.

Predictions: depth decreases from center to edge, that is, the ordering is inverted. All models produce exactly the opposite profile. The mean prediction of DA-V2-Metric decreases monotonically from 30.25 m (center) to 11.80 m (edge), by 18.45 m overall; that of Metric3D decreases from 96.32 m to 24.69 m, by 71.63 m. Moreover, all mean predictions lie far behind the 2.00 m base plane.

Conclusion. The two observations jointly establish reversal rather than overestimation: (i) the center-to-edge ordering of the predicted depth is the mirror image of the geometric truth, and the protruding apex is perceived as the farthest part of the patch; and (ii) the patch as a whole is displaced far beyond its true location. This confirms that these models reverse the depth profile of the physically protruding patch in this example.

Table R3.3. Geometric reference depths and mean predicted depths across the five regions in the larger-scale physical experiment.

Depth (m)Region 1 (center)Region 2Region 3Region 4Region 5 (outermost)
Geometric reference depth of the patch1.21.41.61.82.0
DA-V2-Metric mean prediction30.2522.7219.1114.9611.80
ZoeDepth mean prediction10.389.217.125.294.09
Metric3D mean prediction96.3265.9147.7830.7124.69
AuthorsOpenReview ↗

Response regarding to Weakness 3.

W3. Does the depth deviation persist during approach, and what does this imply for downstream driving behavior?

Thanks for raising this question. Metric3D's depth deviation persisted during the recorded approach video and exceeded 40 m at all five plotted samples spanning the final 2.0 s.

Setup. We recorded a continuous 10-second video (300 frames at 30 fps) and evaluated Metric3D on every third frame, yielding 100 samples. For each sample, we computed EdE_d within the patch region relative to the paired white-masked reference and reported it in meters.

(1) Temporal results. Across the 20 plotted points (samples 5–100, spanning 9.5 s), EdE_d ranged from 1.81 to 46.33 m, with a median of 8.73 m. The deviation was below 10 m at all plotted points through sample 55 and above 15 m at all nine plotted points from samples 60–100, spanning 4.0 s. It peaked at 46.331 m at sample 95 and decreased slightly to 43.659 m at sample 100.

(2) CARLA evaluation. Our study focuses on monocular depth estimation, and our evaluation pipeline ends at the predicted depth maps. A closed-loop CARLA evaluation would require integrating those predictions into a driving stack, which is beyond the scope of this study.

Reviewer_sEfeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Phy-DRILL, a 2.5D physical adversarial attack against monocular depth estimation (MDE) models. By exploiting the Depth Reversal Illusion, the method combines parameterized frustum geometry with optimized surface textures to make physically convex objects appear as concave regions. Experiments across six MDE models demonstrate strong attack effectiveness in both digital and physical environments, including black-box transferability and robustness under varying viewing conditions.

Critical strengths
  1. Novel attack perspective. The paper introduces a geometry-aware physical attack inspired by the Depth Reversal Illusion, moving beyond conventional texture-only adversarial patches. Jointly optimizing geometry and texture provides an interesting direction for physical adversarial attacks.

  2. Comprehensive experimental evaluation. The paper evaluates the proposed method across six MDE models in both digital and physical environments. The results demonstrate strong attack performance, black-box transferability, and robustness under different viewing angles and environmental conditions.

Critical weaknesses
  1. Insufficient quantitative evidence for the claimed depth reversal phenomenon. The central claim of this paper is that Phy-DRILL causes monocular depth estimation models to perceive physically convex structures as concave regions. However, the adversarial objective primarily maximizes the average predicted depth, while the evaluation metrics EdE_d, RaR_a, and AadeA_{ade} measure depth estimation errors rather than geometric reversal. A substantial overestimation of depth does not necessarily imply that the model has reversed the underlying surface geometry. Although qualitative visualizations are provided, the paper lacks direct quantitative evidence of depth reversal, such as an analysis of local depth ordering or predicted surface geometry. This makes it difficult to determine whether the observed effects genuinely arise from the proposed Depth Reversal Illusion or primarily reflect depth overestimation.

  2. Limited physical validation under realistic observation scales and distances. Although the paper demonstrates promising results in physical environments, the main physical comparison is conducted using a small-scale structure with a height of 0.1 m at a camera distance of 0.5 m. While additional experiments examine different viewing angles, environmental conditions, and image occupancy ratios, the effectiveness of the attack at larger physical scales and more practical observation distances remains insufficiently established. Given that physical robustness is a central motivation of this work, further controlled physical experiments with varying object sizes and camera distances would help substantiate the claimed real-world applicability, without requiring deployment in actual autonomous driving systems.

Questions for authors
  1. Could the authors provide direct quantitative evidence of depth reversal, such as local depth ordering or surface geometry analysis, to distinguish it from simple depth overestimation?

  2. Could the authors evaluate the attack at larger physical scales and longer observation distances in controlled environments?

  3. Why is π-Jack excluded from the baselines, given its related perspective-based attack mechanism? Could the authors provide a comparison or explain the differences in attack settings?

  4. Could the authors compare fixed frustum geometry with jointly optimized geometry and texture to clarify the independent contribution of geometry optimization?

6 comments
AuthorsOpenReview ↗

Response regarding to Weakness 1 and Question 1.

What Does Local Depth Ordering Show About Depth Reversal? (W1 / Q1)

Thanks for the questions. To validate that our method truly achieves depth reversal rather than mere overestimation, we conducted an additional physical experiment that measures the spatial profile of the predicted depth across the patch, from its protruding center to its edge.

Setup: We fabricated a larger patch from printed KT foam board and mounted it at the rear of a passenger car. A frontal-view photograph was captured in a real road scene, with the camera 2.0 m from the patch's base plane. Within the visible patch region, we defined five concentric rectangles: one central region plus four non-overlapping bands, so that we can examine how predicted depth changes from the center to the edge. Table R2.1 reports the geometric reference depth and the mean predicted depths of three metric MDE models in each region.

Geometric reference (Real Depth): depth increases from center to edge. Because the patch physically protrudes toward the camera, its geometric reference depth increases monotonically from Region 1 (center, 1.2 m) to Region 5 (outermost band, 2.0 m), by 0.80 m overall; the four inner regions all lie in front of the 2.00 m base plane. Note that a model which merely overestimates depth while still interpreting the surface as convex would preserve this ordering — closest at the center, farthest at the edge.

Predictions: depth decreases from center to edge, that is, the ordering is inverted. All models produce exactly the opposite profile. The mean prediction of DA-V2-Metric decreases monotonically from 30.25 m (center) to 11.80 m (edge), by 18.45 m overall; that of Metric3D decreases from 96.32 m to 24.69 m, by 71.63 m. Moreover, all mean predictions lie far behind the 2.00 m base plane.

Conclusion. The two observations jointly establish reversal rather than overestimation: (i) the center-to-edge ordering of the predicted depth is the mirror image of the geometric truth, and the protruding apex is perceived as the farthest part of the patch; and (ii) the patch as a whole is displaced far beyond its true location. This confirms that these models reverse the depth profile of the physically protruding patch in this example.

Table R2.1. Geometric reference depths and mean predicted depths across the five regions in the larger-scale physical experiment.

Depth (m)Region 1 (center)Region 2Region 3Region 4Region 5 (outermost)
Geometric reference depth of the patch1.21.41.61.82.0
DA-V2-Metric mean prediction30.2522.7219.1114.9611.80
ZoeDepth mean prediction10.389.217.125.294.09
Metric3D mean prediction96.3265.9147.7830.7124.69

We have added this region-wise physical analysis to the revised manuscript as direct quantitative evidence of depth reversal.

AuthorsOpenReview ↗

Response regarding to Weakness 1

How Do the Mean-Depth Objective and Error Metrics Relate to the Depth Reversal Criterion? (W1)

Thank you for the comments. Nevertheless, we note that the comments conflate two complementary parts of our evaluation: quantifying the strength of the depth attack and determining whether the predicted depth profile is reversed. Our objective maximizes the mean predicted depth within the patch, a differentiable surrogate aligned with the attack goal of pushing the patch toward a farther predicted depth. It is not intended to serve as a reversal indicator by itself.

Likewise, EdE_d, RaR_a, and AadeA_{ade} quantify complementary aspects of attack strength: the magnitude of depth deviation, its spatial coverage, and its robustness across viewing angles. These metrics are deliberately continuous performance measures; they do not encode the sign or local ordering of depth predictions. We therefore do not use a large value of any one of these metrics alone to claim geometric reversal.

We determine reversal using a separate, stricter criterion. Since the physical patch protrudes in front of its base plane, a prediction that overestimates depth but remains in front of that plane is not counted as reversal. Reversal requires the predicted patch depth to cross behind the base plane. This reference is validated using a paired image in which the patch texture is removed by filling the patch region with white; the predicted depth of that region agrees with the physically measured base-plane distance. The base-plane criterion is also conservative: once a prediction crosses behind the plane, its deviation from the base-plane reference is smaller than its deviation from the actual protruding surface.

Finally, we directly examine the spatial ordering of the predicted depth in the additional physical experiment (Table R2.1). The geometric depth increases from the patch center to its edge, whereas the regional mean predictions of all three models decrease in the same direction and lie behind the base plane. Thus, our evaluation does not equate depth overestimation with reversal: the aggregate metrics quantify attack strength, while the validated base-plane criterion and region-wise ordering test reversal directly.

AuthorsOpenReview ↗

Response regarding to Weakness 2 and Question 2.

How Does the Attack Perform at a Larger Physical Scale and Longer Observation Distances? (W2 / Q2)

Thank you for specifying the need to evaluate larger physical scales and longer observation distances. We conducted an additional physical study using a patch eight times taller than the 100 mm patch in the original experiment and tested it at distances 4–16 times farther than the original 0.5 m distance.

Setup. The main physical comparison in the original submission used a 100 mm-high patch observed from 0.5 m. To extend the evaluation to a larger physical scale and longer distances, we fabricated an 800 mm-high patch from printed KT foam board and mounted it at the rear of a sedan. The patch has base and top-face edge lengths of 722 mm and 123 mm, respectively. We captured road-scene photographs at seven observation distances, from 2.0 m to 8.0 m in 1.0 m increments, and evaluated the depth predictions offline. We also evaluated viewing angles from −70° to +70° at a fixed observation distance of 4.0 m.

Results. For each photograph, we calculate EdE_d as the mean absolute deviation from the paired reference prediction within the patch mask. As shown in Table R2.2, the deviations generally become smaller at longer distances. For example, DA-V2-Metric's EdE_d decreases steadily from 9.46 at 2.0 m to 1.43 at 8.0 m. Metric3D's EdE_d drops from 43.66 at 2.0 m to 3.40 at 5.0 m and fluctuates at longer distances, reaching 6.14 at 8.0 m. We interpret these changes within each model's output scale.

Conclusion. The distance experiment shows how the magnitude of the prediction deviation varies over 2.0–8.0 m. Table R2.1 provides direct evidence of depth reversal in frontal view, where all models reverse the center-to-edge depth order and place the mean depths of the protruding regions behind the base plane.

Table R2.2. Mean absolute prediction deviation EdE_d at different observation distances.

Observation distance (m)2345678
DA-V2-Metric9.465.815.544.263.522.951.43
ZoeDepth4.242.593.270.473.270.270.08
Metric3D43.6642.925.443.405.861.636.14
Monodepth210.028.848.064.821.860.720.47
Reviewer_DLHHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Phy-DRILL, a physical attack method for monocular depth estimation, inspired by the "depth-reversal illusion." The approach constructs a protruding frustum with perspective-consistent textures and alternately optimizes its geometric parameters and surface appearance. Experiments encompass digital compositing, controlled physical deployment, and transfer from a Depth Anything V2 surrogate to five additional model entries.

Critical strengths
  1. The paper presents a physical adversarial attack against monocular depth estimation by combining protruding geometric structures with conflicting visual depth cues.

  2. The physical design is simple and easy to construct. The experiments cover multiple victim models, component ablations, and additional evaluations under different viewing angles, distances, and environmental conditions.

Critical weaknesses
  1. The adversarial objective does not explicitly enforce depth reversal. Eq (6) only maximizes the average predicted depth within the attacked region, which may lead to depth overestimation without reversing the relative depth ordering.
  2. The paper uses the standard Depth Anything V2 as the white-box surrogate model and evaluates transferability to other models, including Depth Anything V2-Metric, Depth Pro, and Metric3D. However, the standard Depth Anything V2 predicts relative inverse depth, whereas metric depth models predict physical distances.
  3. Approximating the ground-truth depth of an entire 3D object using a single distance may introduce non-negligible errors. The authors should compute pixel-wise ground-truth depth using the known 3D geometry and camera parameters, rather than relying on a single distance approximation.
Questions for authors

Please see Weakness.

  1. Can the authors provide direct evidence of depth reversal rather than depth overestimation?
  2. How are different depth representations handled during optimization and evaluation?
  3. How does using a single ground-truth depth affect the reliability of the physical evaluation?
7 comments
AuthorsOpenReview ↗

Response regarding to Weakness 1

Can Eq. (6) Reverse the Relative Depth Order by Maximizing Mean Predicted Depth? (W1)

Thanks for the questions. We believe this concern reflects a misunderstanding of the role of Eq. (6). In our corridor setting, Eq. (6) is used to achieve a stronger attack effect that combines depth reversal with placing the predicted surface behind the patch's base plane.

(1) The starting point of optimization already leans toward reversal. Our texture is not adversarial noise; it is a perspective-consistent "corridor" texture generated by Stable Diffusion and mapped onto the frustum. Sec. 3.1 shows this structure already flips the interpretation of the surface at the beginning: with 4 vanishing points and a matching texture gradient, DA-V2-Metric predicts 29 m where the ground truth is 10 m (Fig. 4), and the predicted depth maps depict the protruding structure as a receding surface (Fig. 3). In other words, reversal is the natural failure mode of this geometry+texture combination, and Eq. (6) simply amplifies it.

(2) The “corridor” structure makes it possible for maximizing the average depth to result in depth reversal: i. the structure restricts the solution space to "corridor" explanations. MDE models infer depth from perspective cues. A 4-VP perspective-consistent corridor texture admits essentially one family of depth explanations: a corridor receding toward the vanishing points in which predicted depth must increase or decrease monotonically from the corridor entrance (the patch border in the image) to the corridor end (the patch center). Then, maximizing the mean stretches the corridor, leading to increase monotonically. ii. The border is anchored; the center is free. The patch border adjoins real scene pixels, so its predicted depth is anchored by scene continuity and cannot be pushed far. The corridor's end has no such anchor and its depth is unbounded from above.

In summary, since the corridor structure anchors the patch border but leaves the center free and unbounded, maximizing the average depth could only be achieved by pushing the center (i.e., the closest point in reality) to be the farthest, resulting in depth reversal.

AuthorsOpenReview ↗

Response regarding to Question 1

What Direct Evidence Distinguishes Depth Reversal from Depth Overestimation? (Q1)

Thanks for the questions. To validate that our method truly achieves depth reversal rather than mere overestimation, we conducted an additional physical experiment that measures the spatial profile of the predicted depth across the patch, from its protruding center to its edge.

Setup: We fabricated a larger patch from printed KT foam board and mounted it at the rear of a car. The patch is 800 mm high, with base and top-face edge lengths of 722 mm and 123 mm, respectively. A frontal-view photograph was captured in a real road scene, with the camera 2.0 m from the patch's base plane. Within the visible patch region, we defined five concentric rectangles: one central region plus four non-overlapping bands, so that we can examine how predicted depth changes from the center to the edge. Table R1.1 reports the geometric reference depth and the mean predicted depths of three metric MDE models in each region.

Geometric reference (Real Depth): depth increases from center to edge. Because the patch physically protrudes toward the camera, its geometric reference depth increases monotonically from Region 1 (center, 1.2 m) to Region 5 (outermost band, 2.0 m), by 0.80 m overall; the four inner regions all lie in front of the 2.00 m base plane. Note that a model which merely overestimates depth while still interpreting the surface as convex would preserve this ordering — closest at the center, farthest at the edge.

Predictions: depth decreases from center to edge, that is, the ordering is inverted. All models produce exactly the opposite profile. The mean prediction of DA-V2-Metric decreases monotonically from 30.25 m (center) to 11.80 m (edge), by 18.45 m overall; that of Metric3D decreases from 96.32 m to 24.69 m, by 71.63 m. Moreover, all mean predictions lie far behind the 2.00 m base plane.

Conclusion. The two observations jointly establish reversal rather than overestimation: (i) the center-to-edge ordering of the predicted depth is the mirror image of the geometric truth, and the protruding apex is perceived as the farthest part of the patch; and (ii) the patch as a whole is displaced far beyond its true location. This confirms that these models reverse the depth profile of the physically protruding patch in this example.

Table R1.1. Geometric reference depths and mean predicted depths across the five regions in the larger-scale physical experiment.

Depth (m)Region 1 (center)Region 2Region 3Region 4Region 5 (outermost)
Geometric reference depth of the patch1.21.41.61.82.0
DA-V2-Metric mean prediction30.2522.7219.1114.9611.80
ZoeDepth mean prediction10.389.217.125.294.09
Metric3D mean prediction96.3265.9147.7830.7124.69
AuthorsOpenReview ↗

Response regarding to Weakness 2 and Question 2.

How Are Relative Inverse Depth and Metric Depth Handled During Optimization and Evaluation? (W2 / Q2)

Thanks for this question. The two stages involve different sets of models, and we handle each representation at each stage as follows.

(1) During optimization, only the relative representation is involved. The sole model in the optimization loop is the white-box surrogate, standard DA-V2; the other five models including all metric models are attacked in a black-box manner and never enter optimization. Either a relative-depth or a metric-depth model can serve as the surrogate as long as its output can be expressed as a differentiable score that increases with predicted distance, and we chose standard DA-V2 for this study. DA-V2 outputs relative depth q, where a larger q indicates a nearer prediction. We therefore minimize the mean of q inside the patch mask, which is exactly Eq. (6) with f = −q, so that larger f corresponds to a farther prediction. Since the objective only requires this direction of change, the native relative output is used directly, without converting it to meters — no metric representation ever appears in optimization.

(2) During evaluation, the two representations are handled separately: i. Relative-depth models (DA-V2, Monodepth2): the prediction deviations are expressed on each model's own relative scale, with no scale conversion. ii. Metric-depth models (DA-V2-Metric, Depth Pro, Metric3D, ZoeDepth): each model produces metric depth on its own, that is, DA-V2-Metric and ZoeDepth output depth in meters directly; Depth Pro infers the focal length from the image and recovers metric depth from it; Metric3D predicts depth in canonical camera space, converted to meters using the focal length of the input image. The deviations are then computed in meters.

Because output scales and calibration procedures differ across models, absolute deviation values are not comparable across models, but our protocol never requires them to be: every comparison in the paper is between different attack methods under the same victim model, so representation differences do not affect any conclusion. This also holds for the transfer experiments: the surrogate is used only during optimization, and each black-box victim is evaluated in its own representation.

We have clarified both stages in Sec. 4.1 of the revised manuscript. In addition, we chose standard DA-V2 as the surrogate model, because it provides the relative depth signal required by Eq. (6) with lower costs (See Table R1.2).

Table R1.2. Computational costs of different MDE models as surrogates for optimization.

Model configurationParameters (M)Reference encoder inputEncoder forward cost (GMAC)
DA-V224.79518 × 51846.39
DA-V2-Metric24.79518 × 51846.39
Metric3D37.50616 × 1064174.48
ZoeDepth344.82384 × 768413.54
Depth Pro951.99Full image7052.65

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.