Person Reconstruction — People from Incomplete Evidence
Introduction
Person Reconstruction rebuilds a person (face, build, posture) in 3D from evidence that does not show them directly. The input is incomplete: a witness's memory, photographs of a different time, or remains. The model is an argument made visible. It carries only what the evidence supports, and it labels each feature that is an inference.
Situated Testimony works with the person present, walking the analyst through a finished scene. Person reconstruction works with the person absent: the analyst walks the witness through the model instead.
From Memory
A witness describes a person. The model iterates toward their recollection. The Bilthoven cold case built a suspect this way.
From Photographs
Photos from different moments (life, crime scene, autopsy) are reconciled into one 3D person. The USF homicide case did this for three victims.
From Capture
A 26-camera rig captures a present person as a true-scale 3D face for surveillance comparison. The Zurich system (2019) makes this routine.
1. The Input Channels
Each channel carries a different kind of uncertainty. Name the channel on every model. A reviewer must be able to answer: what was observed, and what was inferred?
| Channel | Input | What the model may claim | Typical failure |
|---|---|---|---|
| Memory | Witness description, composite sketch | Plausible person consistent with the account | The modeler's face replaces the witness's person |
| Photographs | Multiple photos across time | A specific person at a specific moment | Features averaged from incompatible views |
| Capture | Photogrammetry of the present person | A measured person | Angle mismatch with the footage being compared |
The first two are reconstruction (incomplete evidence to 3D). The third is capture (person to 3D). Capture is standardized and mature. Reconstruction has no standard workflow, and the defensibility work happens in that gap.
2. The Iterative Loop
The Bilthoven pipeline is the reference shape. It has four stages, and you run them as a loop.
Stage 1 — Composite (2D)
A composite artist works with the witness, feature by feature: eyebrows, hairline, face shape, build. The witness corrects. The 2D result is the shared reference image for every later stage.
Stage 2 — First 3D Pass
A 3D artist converts the composite into a face mesh. The priority is face form, then hair length, curls, and clothing descriptions. This pass produces a hypothesis.
Stage 3 — Witness Feedback Sessions
Multiple sessions. The witness reviews the model and names what is wrong: face shape, hair, build, posture, height. Each correction is baked into the model and logged. The model refines toward the whole person. People recognize build and posture as much as the face.
Stage 4 — Release with a Verification Path
In Bilthoven, the model went public only after the perpetrator failed to come forward, and every tip had a DNA check behind it. The model converts recognition into identification. The DNA check is what makes that conversion hold.
Session log, every round:
Session 3 — 2026-xx-xx
Witness changes: fuller build, shorter hair, left shoulder higher
Model version: v3 → v4
Not changed (witness confirmed): eye color, face width
Open: scar on chin — witness uncertain, keep both variants
3. The Method Behind the Plausibility
The science is a 3D morphable face model: a statistical model of face geometry learned from a database of scans. Blanz and co-authors published the witness-driven method in 2006. Three mechanisms matter for the analyst:
- Attribute constraints. The witness's description (hair, gender, age, build, ethnicity) restricts face space to a residual subspace. Only faces matching the constraints remain possible.
- Completion from learned correlations. Unspecified features are completed from anatomical correlations, not invented. The model fills gaps with what is statistically plausible for that subspace.
- Adaptation to the local population. The database can be tuned to the population the person likely belongs to. A generic database biases the result.
The 2006 evaluation compared the system against a professional forensic artist working with the standard police composite tool. The comparison is not a claim of superiority. A 3D-from-memory model is a working instrument: it structures recall and tests consistency, and it does not identify on its own.
4. The Defensibility Record
In USA v. Astarita (D. Or., 2018), the court excluded the human figures from a 3D scene reconstruction. The judge held that the figures' positions "could not be reasonably inferred" from the blurry video, and that showing them implied an unjustified certainty (Federal Rule of Evidence 403, undue prejudice). The court did allow the 3D scene as context for an eyewitness to describe where the figures stood.
The court drew a line: the 3D scene supports an eyewitness, but the figures themselves carry no testimony.
Rules that keep a person-reconstruction model on the right side of that line:
- State the input channel on the model and in every write-up. If a model mixes memory and photos, label the mix.
- Version every witness round. The log in Stage 3 is evidence. It shows the model followed the witness, not the modeler.
- Separate observed from inferred. Every feature that no source showed is an inference. List them. A variant list (scar: yes/no) is stronger than a silent choice.
- Do not let the model override the witness. If the witness says "no" to a feature the statistics suggest, the witness wins.
- Record the sessions. Screen-record the review sessions the way Situated Testimony records its live sessions. The performance of navigating is part of the evidence chain.
- Match the claim to the evidence. The model may show a person consistent with the account. It may not show the person.
5. Where the Blender Work Happens
The reconstruction itself is standard modeling work, and the reference layer already covers it:
- Sculpting — face form from the composite.
- Direct Modeling — hair, clothing, build.
- Measurements, Scale & Dimension — build, height, and posture are measurements. Document the scale basis for each.
Summary
Person reconstruction produces a visible hypothesis from incomplete evidence. The loop runs four stages: composite, first pass, witness sessions, release with a verification path. The record is what makes the model hold up in review, and it must answer three questions: which channel drove each feature, what the witness confirmed, and what remains an inference.
The core rule: build the model from the witness and the evidence, and do not present it as a substitute for either.
Resources
- Bilthoven Cold-Case 3D Suspect Model — the case study this technique is documented from
- Situated Testimony — the present-witness direction of the same technique class
- Sculpting — the face-form work
Sources
- Politie toont 3D-model in verkrachtingszaak | Bilthoven — the Opsporing Verzocht film, and the Politie news item
- USF — High-Profile Homicide Trial Showcases Evidence Recreated at USF (2020) — 3D victim skulls from photographs, admitted as evidence; Kaplan expert testimony video
- Judge throws out 3D model — fatpencilstudio — USA v. Astarita, D. Or. (2018), the Rule 403 exclusion of the human figures
- Making 3D Mugshots — Forensic Magazine — the Zurich 26-camera capture system, Forensic Science International (2019)
- Blanz, Albrecht, Haber, Seidel — Creating Face Models from Vague Mental Images, Computer Graphics Forum 25(3), 2006 — the morphable-model method