Software that turns a single photo of a person into a 3D body model has to guess. A photo flattens depth, hides the back of the body, and hides any limb behind another limb. Modern systems fill those gaps with statistical expectations of what a "typical" body looks like, learned from a few dozen to a few thousand scanned people. When the person in the photo differs from those expectations, the system does not report uncertainty. It quietly returns a body closer to the average.
This makes monocular 3D reconstruction a good setting for teaching engineering students a lesson that ethics curricula find hard to teach with real tools: some failures are limits on what the input can tell you, and no amount of tuning removes them. In this setting the responsible engineering move is sometimes to refuse to answer. We propose a hands-on lab in which students measure how much of a reconstruction comes from the photo and how much comes from the learned prior. They then write and test a rule that tells the system when to abstain. We will evaluate the lab with a code-graded outcome rather than attitude surveys.
Core Claim
Integrated ethics education has produced convincing hands-on curricula in language processing and tabular machine learning. It has not produced them in geometric vision. Its most rigorous studies also measure learning mainly through surveys, which hit a ceiling because students already say they care about ethics. Our claim is narrow and testable. Students learn to audit and constrain a 3D system more effectively when three things happen together: they quantify, on real benchmark data, how far the system's output drifts toward the average body as image evidence disappears; they compare single-view and multi-view reconstructions of the same subject; and they must ship an abstention rule that is scored automatically. Students who only run a disaggregated error audit, or who receive the same ethics content as a separate lecture with a documentation assignment, should do worse.
The new element is not "ethics in a 3D vision course." The new element is that the "No Technical Fix" realization becomes a measured quantity and a graded artifact. A student's ethical judgment about when not to predict gets scored by whether it actually lowers error, and lowers the gap between subgroups, on held-out data.
Where This Sits
Three lines of education work motivate the design. Each stops short of geometric vision. Integrating Ethics into the NLP Curriculum builds exercises on dual use, bias, and privacy for text models. We borrow its stance that ethics should be a technical competency, but its exercises have no analogue for depth ambiguity or body-shape priors. Authentic Integration of Ethics and AI through Sociotechnical, Problem-Based Learning introduces the 4D framework and documents students reaching the "No Technical Fix" realization on synthetic tabular data. It treats deciding not to build as a valid engineering outcome. We take that "Determination" lens directly. Its evidence is qualitative and comes from two small cohorts, and we replace that with a scored outcome. Computational Thinking with Computer Vision shows that vision is an engaging hook in an introductory course and reports significant self-efficacy gains. It stays in 2D and explicitly runs into survey ceiling effects. That ceiling is the measurement problem our design is built to avoid.
Two further papers supply the method. Reproducibility as a Mechanism for Teaching FACT in AI shows that re-running published models on single GPUs makes fairness concrete: 30 of 32 student groups submitted reports to the reproducibility challenge, and 9 were published. Exploring Social Biases of Large Language Models in a College AI Course has students build their own bias probes. We reuse both moves, applying them to body-reconstruction pipelines instead of classifiers and text models.
Understanding ML Practitioners' Data Documentation Perceptions constrains the design. It finds that practitioners treat datasheets as a box-checking exercise. That is the reason the documentation-only condition is our naive reference arm and not our intervention.
On the vision side, the technical substrate comes from four papers.
- Keep It SMPL fits a parametric body model to detected 2D joints under explicit pose and shape priors and a gender-neutral template. Students can change the weight on the prior and watch the effect directly.
- DeforHMR is a feed-forward transformer that regresses the same body parameters. Its own reported numbers degrade out of distribution, and it concedes failures under severe occlusion.
- INT is a second strong regressor that tokenizes joints and shape.
- PKU-DyMVHumans is a 56–60 camera capture of 32 subjects, including 5 children. It makes the single-view versus multi-view comparison possible for the same person.
FFMN applies single-view body fitting to monitoring live-line electrical workers. It gives a realistic deployment scenario in which a confident wrong reconstruction has consequences.
Research Program
We will build an open lab module of about three weeks that plugs into an existing upper-level vision or machine learning course. It has four parts.
Part 1: measure the prior's pull. Students run SMPLify-style fitting and a pretrained regressor (DeforHMR or INT) on 3DPW, RICH, and PKU-DyMVHumans frames. For each subject they compute a shape-shrinkage ratio,
$$ s = \frac{\lVert \hat\beta - \bar\beta \rVert}{\lVert \beta^{\text{ref}} - \bar\beta \rVert}, $$
where $\hat\beta$ is the single-view estimate of the body-shape parameters, $\beta^{\text{ref}}$ is a reference shape, and $\bar\beta$ is the mean shape of the model. The reference comes from the provided ground truth, or from a multi-view fit on PKU-DyMVHumans. A value of $s$ well below 1 means the system pulled the person toward the average body. Students track $s$ as they raise the weight on the shape prior, as they crop or occlude limbs, and across subgroups the data can support. Those subgroups are children versus adults in PKU-DyMVHumans, plus body-shape quantiles of the reference shapes. We will not report $s$ for ethnicity or other attributes the datasets do not label.
Part 2: measure what one view loses. On PKU-DyMVHumans, students reconstruct each subject from one camera and from many cameras. The gap between the two is information the single photo lacks. Students then try technical fixes, such as different models, prior weights, and test-time augmentation. The question is whether any of them close the gap for the subjects who were pulled hardest toward the average.
Part 3: write an abstention rule. Students must write a rule that decides when the system should refuse to output a body. The rule may use only information available at test time: 2D keypoint confidence, disagreement between the two regressors, residuals from the fitting process, or the estimated $s$-proxy. The deployment case is FFMN-style worker monitoring. Each rule is scored automatically on held-out RICH and occluded 3DPW frames. The score is the error on accepted predictions at a fixed coverage $c$ (the fraction of inputs the rule accepts), together with the largest gap in that error between subgroups:
$$ \text{Score}(c) = \text{MPJPE}_{\text{accepted}}(c) + \gamma \max_{g,g'} \left| \text{MPJPE}_g(c) - \text{MPJPE}_{g'}(c) \right|. $$
Here MPJPE is mean per-joint position error in millimeters, $g, g'$ range over subgroups, and $\gamma$ is a fixed weight announced in advance.
Part 4: argue the determination case. Students write a short determination memo in the 4D style. It argues whether the monitoring system should be deployed at all, at what coverage, and with what consent arrangements.
The research output has three pieces: the released lab, an auditing toolkit, and a pre-registered comparison of teaching conditions. A small empirical by-product is a set of disaggregated shrinkage and error measurements on these benchmarks. To our knowledge from the file, nothing like it has been reported for these models.
Baselines
The study compares teaching conditions, so the baselines are alternative lessons. All arms share the same instructor time (three weeks), the same GPU allocation per student, and the same pretrained models. All take the same final held-out assessment, graded blind to condition.
The naive reference is a standalone ethics lecture on 3D avatars: dual use, consent, and dataset demographics, plus a datasheet assignment for the training data. Students run the same reconstruction code without any auditing task. This arm represents the common practice of bolting ethics onto a course. Heger et al. predict it will produce box-checking. Our arm beats it if students produce abstention rules that score better on $\text{Score}(c)$, and if their memos correctly attribute failures to missing information versus model capacity more often.
The strongest existing pedagogy is the Krakowski 4D problem-based module run as published, on its tabular admissions and hiring data. It has the same hours and is followed by the same 3D assessment. This is the strongest competitor because it is the only cited curriculum shown to produce the "No Technical Fix" realization, and porting it is cheap. For this arm, no reported quantitative learning result was found on Lacuna; its evidence is qualitative. Beating it means a higher score on the 3D assessment. We will also test transfer in the other direction, on a held-out tabular audit, to check that our arm does not lose general skill.
The ablation that isolates the key ingredient is a technical-only audit. Students run the same disaggregated MPJPE and per-vertex error analysis on the same data and receive the same toolkit. They do not do the single-view versus multi-view comparison, are not asked to attribute failures to missing information, and receive no determination framing. They still submit an abstention rule, framed as an accuracy optimization. If this arm matches the full lab, the ethics framing adds nothing beyond good engineering practice, and the claim fails.
For the vision systems inside the lab, the reference points students must reproduce are numbers the cited papers report. DeforHMR reports 63.6 mm MPJPE and 75.2 mm per-vertex error on 3DPW, and 84.2 mm MPJPE on RICH. INT reports 42.0 mm PA-MPJPE on 3DPW. SMPLify reports 79.9 mm mean joint error on HumanEva-I. Students who cannot reproduce DeforHMR's figures within a stated tolerance are flagged, so that a bad baseline cannot inflate their abstention scores. The no-abstention operating point (coverage 1) is the floor every student rule must beat.
Why not just…?
Why not just keep ethics in a separate elective? A separate course cannot score an engineering decision against real model behavior, and scoring that decision is the outcome we care about. The material on superficial layering and Heger's documentation findings suggest separated content stays abstract. That is a weak argument, not proof. The naive-reference arm approximates the elective inside the same course, so the experiment tests this directly. If that arm matches ours on the code-graded assessment, the integration claim is refuted.
Why not just reuse the tabular or NLP modules that already work? This is a real threat. The Krakowski module is the only one with documented "No Technical Fix" outcomes, and the underlying lesson may transfer. Our argument is that tabular data has no analogue of projection destroying depth, and none of a shape prior pulling bodies toward the average. Students taught on tabular data may recognize bias without being able to identify missing information in a geometric pipeline. The ported-4D arm settles this. If it equals ours on the 3D assessment, the right recommendation is to reuse the existing module and drop ours.
Why not just require datasheets and model cards? Heger et al. find practitioners treat these as a box-checking exercise. More concretely, a datasheet records who is in PKU-DyMVHumans but does not reveal what the model does to children's bodies. The naive arm includes a datasheet, so its contribution is measured rather than assumed.
Why not just teach a rigorous disaggregated audit and skip the ethics framing? This is the most serious alternative. Good engineers already report per-group error, and Lučić et al. suggest hands-on re-implementation does most of the work. The technical-only ablation exists for this reason. We predict the difference shows up specifically in abstention rules on the most occluded and most out-of-distribution subjects. Students who have not compared single-view against multi-view reconstructions will chase accuracy with model changes rather than abstaining. If that prediction fails, we will report the integrated framing as unnecessary.
Why not just wait for larger diffusion priors to fix the failures? PKU-DyMVHumans reports that even dense multi-view capture with strong neural surface methods fails in concave regions and on changing topology. DeforHMR degrades from 63.6 mm on 3DPW to 84.2 mm on RICH. Part 2 of the lab tests generative fixes empirically by asking whether any available method closes the single-view versus multi-view gap for the most-shrunk subjects. If one does, students learn that too, and the abstention lesson narrows.
Proposed Experiments
Each teaching arm is a lab section, and sections are randomized across at least two course offerings at two institutions. The analysis is pre-registered and uses mixed-effects models with section as a random effect. Small cohorts are a known limitation of this literature, so we will report a power analysis on the observed intra-section correlation and treat a single offering as a pilot.
The primary outcome is $\text{Score}(c)$ at coverage 0.8 on a held-out evaluation pool. That pool is RICH frames plus synthetically occluded 3DPW frames that no student has seen, and it is computed automatically. Secondary outcomes come from a blind rubric applied to the determination memo. The rubric checks whether the student correctly separates failures caused by missing information from failures a better model could fix, whether they propose coverage limits, and whether they raise consent and dual-use issues for single-photo avatars of the kind enabled by GAF and PIFu. The pre- and post-course ethics-concern survey from the Chowdhury setting is included mainly to show whether it saturates while the scored outcome still separates the arms.
First decisive check. The full lab must beat the technical-only ablation on the subgroup-gap term of $\text{Score}(c)$, not only on overall error. If it does not, the ethics framing is not doing what we claim.
Second decisive check. The full lab must beat the ported-4D arm on the 3D assessment without losing to it on the held-out tabular audit. This tests whether the domain matters.
Third check, on the substrate. Students' shrinkage measurements must show $s$ systematically below 1 for children in PKU-DyMVHumans and for extreme body-shape quantiles. If the prior does not pull these subjects toward the average, the premise of the lab is wrong for current models, and we will redesign Part 1 around occlusion alone.
References
- Integrating Ethics into the NLP Curriculum
- Authentic Integration of Ethics and AI through Sociotechnical, Problem-Based Learning
- Computational Thinking with Computer Vision: Developing AI Competency in an Introductory Computer Science Course
- Reproducibility as a Mechanism for Teaching Fairness, Accountability, Confidentiality, and Transparency in Artificial Intelligence
- Exploring Social Biases of Large Language Models in a College Artificial Intelligence Course
- Understanding Machine Learning Practitioners' Data Documentation Perceptions, Needs, Challenges, and Desiderata
- Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image
- DeforHMR: Vision Transformer with Deformable Cross-Attention for 3D Human Mesh Recovery
- Capturing the Motion of Every Joint: 3D Human Pose and Shape Estimation with Independent Tokens
- PKU-DyMVHumans: A Multi-View Video Benchmark for High-Fidelity Dynamic Human Modeling
- FFMN: Fast Fitting Mesh Network for Monocular 3D Human Reconstruction in Live-Line Work Scenarios
- GAF: Gaussian Avatar Reconstruction from Monocular Videos via Multi-view Diffusion
- PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization