When a protective oxide scale on a hot-section alloy fails, it flakes off ("spalls") and leaves shallow pits. The depth and volume of those pits tell an engineer how much protection is left. Today that depth is measured destructively or with lab instruments. In the field, the usual tool is a borescope, a single small camera pushed inside the engine, and a single image cannot separate a dark stain from a dark hole. We propose a single-image method that recovers pit depth in metric units. It does this by first identifying which oxide layer is exposed at each surface point, and only then estimating the remaining fine geometry.
The core observation comes from oxidation science. Mature scales are stratified. On Co-Re-Cr-Si alloys there is an outer CoO layer, an intermediate CoCr₂O₄ spinel, and an inner Cr₂O₃ layer (Oxidation Behavior of Quaternary Co-20Re-25Cr-3Si Alloy at Laboratory Air). Alumina formers grow a continuous Al₂O₃ scale over the metal (Oxidation Behavior of Pre-Strained Polycrystalline Ni3Al-Based Superalloy). Because the scale fails layer by layer, the visible phase at a point largely fixes how deep that point sits: exposed spinel means the outer layer is gone, and bare metal means the whole scale is gone. A material recognizer is therefore also a coarse depth estimator, and the layer thicknesses act as a built-in ruler. That ruler addresses the scale ambiguity that defeats ordinary monocular depth estimation.
Core Claim
Depth on a degraded oxide surface should be predicted as a structured sum. The first part is discrete and comes from recognizing the exposed layer. The second part is a small continuous residual that comes from geometry. The two parts should be computed by separate branches with a gradient barrier between them. Write $d(p)$ for the depth of surface point $p$ below the nominal, unoxidized-scale surface:
$$ d(p) = \sum_{k=1}^{K} \Pr[\ell(p) \ge k]\, t_k \;+\; r(p). $$
Here $\ell(p)$ is the index of the exposed layer, ordered from the outermost layer (0) to bare metal ($K$). $t_k$ is the thickness of layer $k$, taken per specimen from a cross-section or from a parabolic growth law. $r(p)$ is a bounded residual that covers partial spallation, crack walls, raised blisters, and pit-edge slopes.
The layer probabilities come from a recognition branch built to ignore exact position and lighting. The residual comes from a geometry branch built to be position-exact. That second branch samples image features at the projected vertices of a mesh of the part's known nominal surface. We claim this structure removes the stain-versus-pit confusion that defeats off-the-shelf monocular depth models. We also claim it gives metric pit volumes from a single calibrated image.
Relation to Prior Work
The method borrows from three lines of work.
From mesh-aligned sampling we reuse the core operation: project each 3D vertex into image feature maps and read features at that location. Mesh R-CNN does this from a voxel initialization on furniture categories. PyMAF does it in an iterative coarse-to-fine loop on the SMPL human body template. PIFu does it for implicit surfaces of clothed people. All three need a shape prior: a body model, CAD-category statistics, or scans. Our prior is different and stronger. The part's nominal CAD surface is known exactly, so the mesh only has to learn a displacement along the surface normal. These papers also share a weakness: they cannot recover absolute scale from one view. We supply scale through layer thicknesses rather than through a body model.
From head decoupling we take the reason for separating recognition from localization. Several papers motivate the split by showing that one shared head degrades both tasks. The Asymmetric Decoupled head (Object Detection for UAV Aerial Scenarios Based on Vectorized IOU) is credited with part of a VisDrone gain from 34.9% to 44.6% [email protected]. PSOL (Rethinking the Route Towards Weakly Supervised Object Localization) shows that coupled training biases localization toward distinctive parts. DeCo-DETR supplies the specific mechanism we reuse: stop-gradient isolation between the localization and semantic branches. The difference in our setting is how the branches relate. In those detectors the two branches produce unrelated outputs. Here they are recombined through the depth equation above. The open question is whether isolation still helps when the outputs are physically linked.
From monocular ambiguity, the main motivation is Towards Ambiguity-Free Spatial Foundation Model. That paper shows that large depth models commit to a single surface ("converged bias") and coaxes out a second hypothesis with Laplacian prompting of a frozen model. It works on glass, outputs 2D depth layers, and has no notion of material. Our layers are physical strata, and we choose among them by recognizing material. The defect-inspection dataset paper (A dataset for surface defect detection on complex structured parts based on photometric stereo) both constrains and motivates our design. It confirms that stains and dents look alike in a single image, and it resolves the confusion with a strobed multi-light rig over 0.96 s. We use that kind of rig only to create training labels, never at deployment. Finally, the planar-superpixel model of Learning 3-D Scene Structure from a Single Still Image is an older example of putting structure into monocular depth. Its planar facets are too coarse for micrometre-scale relief.
Research Program
Data. We will build a paired dataset, since no such dataset appears to exist. We will oxidize Ni₃Al-based superalloy coupons cyclically at 1100 °C, following the protocol of the pre-strain study. We will include 0%, 5%, and 25% pre-strain so that spallation severity varies. We will also oxidize Co-20Re-25Cr-3Si coupons isothermally at 1000–1200 °C, which gives a multilayer scale. The same coupons will additionally be imaged on curved fixtures to mimic blade leading edges.
Each region is imaged three ways:
- with a calibrated monocular borescope-style camera under one ring light (the deployment condition);
- with a multi-light photometric-stereo rig;
- with white-light profilometry, which gives micrometre-level ground-truth depth.
Phase labels come from SEM/EDS or Raman maps registered to the images. Layer thicknesses $t_k$ come from one cross-section per specimen. A held-out variant estimates $t_k$ instead from exposure time and temperature, using parabolic kinetics; the Co-Re paper reports an activation energy of 220.74 kJ/mol.
We will also build a deliberately adversarial subset. It contains unspalled coupons with heat tint, deposited contaminant, and surface roughness that change appearance without removing any material.
Model. A shared backbone produces a feature pyramid that feeds two branches.
The recognition branch uses multi-scale kernels and pooling, in the spirit of the AD head. It predicts ordinal layer probabilities per vertex, trained with an ordinal cross-entropy loss against the phase maps.
The geometry branch starts from the nominal mesh. It projects the vertices into the pyramid, samples features bilinearly, and runs three coarse-to-fine graph-convolution refinement steps in the style of PyMAF to predict $r(p)$. The residual is bounded to a fraction of the local scale thickness.
The depth loss from profilometry is applied to the summed depth $d(p)$. Following DeCo-DETR, a stop-gradient keeps that loss from flowing into the recognition branch's parameters. As a result, the layer probabilities are learned only from phase labels. Without this barrier, the geometry loss could teach the recognition branch to call stains "spinel" whenever the residual cannot explain them.
Analysis. We will measure how much of the depth variance the discrete term explains compared with the residual, and how this changes with spallation severity. We will also test whether the ruler generalizes: does metric pit-volume error stay low when $t_k$ comes from the kinetic law rather than from a cross-section?
Baselines
The naive reference point is that no pit exists. It predicts zero depth everywhere. This sets the floor for depth error. It also gives a perfect stain false-positive rate, which is why every method must be reported on both pit-depth error and the adversarial subset.
The strongest existing competitor is Depth Anything v2-Large fine-tuned on our profilometry labels. It is the foundation model studied in the converged-bias paper, adapted with a scale-and-shift-aligned depth loss and given the same images, epochs, and hyperparameter search budget as our model. It is strongest because large pretrained depth models are the obvious default and fine-tuning gives them our supervision. Since it predicts relative depth, we give it a per-image least-squares scale and shift fitted against ground truth. This is an oracle advantage our method does not receive. We also evaluate it zero-shot with Laplacian Visual Prompting, taking whichever of its two hypotheses is better. The only reported number for that method is on glass: 75.5% ML-SRA on MD-3k, from Towards Ambiguity-Free Spatial Foundation Model. No reported result was found on Lacuna for any monocular method on oxide-spallation depth. We count it as beating this baseline if three conditions hold over three seeds:
- metric pit-volume error falls by at least 30% relative;
- the stain false-positive rate falls by at least half;
- neither improvement depends on the oracle scale.
The hardware reference is photometric stereo with the YOLOv8-l + TSCM pipeline from the MSDD paper. That paper reports 86.1% mAP on MSDD, on a different task (2D defect detection). It serves as an upper reference that a monocular method should approach on detecting pit versus no pit, not as a target to beat.
Two ablations isolate the key ingredients. The first is a shared-head model that predicts depth directly, with the same backbone and parameter count. It isolates the effect of decoupling. The second is our model with the stop-gradient removed, which isolates the gradient barrier. We will also run a dense depth-map decoder in place of the mesh to isolate the mesh-aligned representation.
Why not just…?
Why not just fine-tune a depth foundation model on profilometry labels? This is the real threat, and it is the main baseline. The converged-bias evidence suggests such models commit to one surface explanation. The MSDD paper reports that stains and dents look the same in a single image. Supervision cannot add information that is missing from the pixels; it can only teach the model to use the material cue. Our hypothesis is that an explicit layer variable uses that cue more reliably than a model that must learn it implicitly. The adversarial heat-tint subset settles the question. If the fine-tuned model matches our stain false-positive rate and volume error there, the structured decomposition is unnecessary and we will say so.
Why not just segment the exposed phase and multiply by layer thickness, with no geometry network? This baseline tests whether the residual earns its place. It should do well on clean, fully spalled pits. It should fail on pit walls, partial spallation, cracks, and blisters. Blisters are raised, and segmentation-only assigns them zero depth. We will run it with the same recognition branch. If its volume error is within 10% of ours, the geometry branch will be dropped from the claim.
Why not just use a per-pixel depth map instead of a mesh? On flat coupons the two representations are equivalent, and we expect the ablation to show no difference there. The mesh matters on curved fixtures and under oblique borescope views. There, the displacement must be measured along the true surface normal so it can be compared with layer thicknesses. Pit volume must likewise be integrated against the nominal surface rather than the image plane. If the dense decoder matches the mesh on curved fixtures too, we will report the mesh as a convenience rather than a contribution.
Why not just use photometric stereo? It works, and we use it for labels. It needs a multi-light rig strobed over about a second. That rig does not fit through a borescope port, and it is impractical on an assembled engine.
Decisive Experiments
Three experiments decide the claim.
The first is the adversarial subset. A method that predicts depth larger than half the outer-layer thickness on unspalled, discolored regions is fooled by stains. We expect the decoupled model to have markedly fewer such points than both the fine-tuned foundation model and the shared-head ablation.
The second is cross-alloy transfer. We train on alumina-forming Ni₃Al and test on the three-layer Co-Re-Cr-Si scale with $t_k$ supplied at test time. This checks whether the ruler formulation carries over when the layer sequence changes, which a model that maps appearance directly to depth cannot do.
The third is a physics consistency check. Predicted spalled-area fraction and volume should rank the 0%, 5%, and 25% pre-strain coupons in the same order as their reported 200-hour weight gains: 0.6944 g/cm² at 5% below 0.9319 g/cm² at 0% below 0.9764 g/cm² at 25% (Oxidation Behavior of Pre-Strained Polycrystalline Ni3Al-Based Superalloy). That would give a nondestructive single-image proxy that agrees with gravimetry.
Supporting metrics are:
- mean absolute depth error inside pits, in µm;
- pit volume error;
- phase mIoU;
- the Chamfer distance of the refined mesh against the profilometry surface.
References
- Oxidation Behavior of Quaternary Co-20Re-25Cr-3Si Alloy at Laboratory Air
- Oxidation Behavior of Pre-Strained Polycrystalline Ni3Al-Based Superalloy
- Mesh R-CNN
- PyMAF: 3D Human Pose and Shape Regression with Pyramidal Mesh Alignment Feedback Loop
- PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization
- Object Detection for UAV Aerial Scenarios Based on Vectorized IOU
- Rethinking the Route Towards Weakly Supervised Object Localization
- DeCo-DETR: Decoupled Cognition DETR for efficient Open-Vocabulary Object Detection
- Multi-Hypothesis Spatial Foundation Model: Rethinking and Decoupling Depth Ambiguity via Laplacian Visual Prompting
- A dataset for surface defect detection on complex structured parts based on photometric stereo
- Learning 3-D Scene Structure from a Single Still Image