Reverse Engineering Explained: A Student Guide - ByteMuni

Reverse engineering turns a physical object into a digital model you can measure, edit and manufacture. In industry, that usually means capturing a part as millions of 3-D points, cleaning those points, turning them into a mesh, and then fitting mathematical surfaces that a CAD system can use. This article is a student-level deep dive into how that pipeline works: the physics of the scanners, the geometry behind the software, the accuracy traps that catch beginners, real case studies, and the legal lines you should know before you start.

It's built around Reverse Engineering: An Industrial Perspective (Springer, 2008), edited by Vinesh Raja of the University of Warwick and Kiran J. Fernandes of the University of York. It's one of the few books dedicated to mechanical reverse engineering from an industry point of view. The book is from 2008, so I've flagged where technology or law has moved on. The fundamentals haven't changed, and they're what this article focuses on.

A note before we start. Most of my career has been in software, where "reverse engineering" means disassemblers and decompilers. This book's first chapter tells a story that sums up the confusion: a factory manager wishes he could photocopy a mechanical part the way a secretary copies a document. When the authors suggest reverse engineering, he assumes it's something to do with programming computers. It isn't, or at least not only. Understanding both meanings will make you a better engineer.

What is reverse engineering?

Engineering normally runs forward: you start with an idea, refine it into drawings and a CAD model, and then manufacture a physical product. Reverse engineering (RE) runs the other way. You start with a physical object that lacks drawings, a bill of materials or a computer model, and you work backward to a digital description of it.

The book offers a few definitions from the research literature, and they all share one idea: producing a CAD model from 3-D points captured by scanning or digitizing an existing part. The chapter by David Page, Andreas Koschan and Mongi Abidi (University of Tennessee) calls the automated version computer-aided reverse engineering (CARE). They describe it as a "3-D fax": a fax machine turns paper into a digital signal, and a CARE system turns a solid object into a CAD model.

Three meanings of "reverse engineering"

Students often mix these up, so it helps to separate them:

• Geometric RE: capturing an object's shape as a CAD model. This is the book's main subject.
• Systems-level RE: taking a machine apart to understand how components interact, or what materials they're made of.
• Software RE: recovering how a program works from its machine code, without the original source.

The Tennessee chapter is explicit that its definition is about geometry and shape, not function. Keep that distinction in mind. A perfect 3-D scan of a gearbox tells you its shape, not its gear ratios under load or the heat treatment of its steel.

Why do companies use reverse engineering?

The book lists many motivations. They fall into a few groups:

• Missing data. The original manufacturer has closed, the product is obsolete, or the documentation was lost or never existed. A classic example is spare parts for aircraft that have been in service for decades.
• Quality and inspection. Comparing a manufactured part against its CAD design, often described as comparing the "as designed" part with the "as manufactured" one.
• Design improvement. Studying wear patterns to fix weak features, strengthening features that have proved themselves in long-term use, and analyzing competitors' products.
• Organic shapes. Car stylists still sculpt in clay, foam or wood. Digitizing the sculpture is often faster and more faithful than trying to recreate it in CAD.
• Personalization. Dental appliances, hearing aids, prosthetics and surgical planning all rely on capturing one specific human body.
• Beyond manufacturing. Film and game assets, art reproduction, architecture, clothing fit and crime scene documentation.

The thread connecting all of these is time. The book frames RE, alongside rapid prototyping, as a way for manufacturers to shorten product development cycles and compete with lower-wage economies through speed and customization rather than cost.

What are the three phases of the reverse engineering process?

Chapter 1 sets out a generic process with three phases:

1. Scanning. Choose a technique, prepare the part, and capture its geometry as a point cloud, a set of (x, y, z) coordinates on the surface.
2. Point processing. Import the points, remove noise, reduce redundant points, and merge multiple scans into one clean cloud.
3. Application-specific model development. Turn the cleaned points into whatever the job needs: a mesh, a CAD surface model, CNC toolpaths, or an inspection report.

Before scanning anything, the book says your strategy should account for the part and the purpose: why you're doing RE at all, how many parts you need to scan, part size and complexity, material (hard or soft), surface finish (shiny or dull), geometry (organic or prismatic, internal or external), and the accuracy you actually need.

That checklist does a lot of work. A shiny, soft, internal-cavity part needs a completely different approach from a dull, rigid, external one. Most beginner failures come from choosing a scanner before answering those questions.

How does 3-D scanning hardware work?

Chapter 3, by D.T. Pham and L.C. Hieu of Cardiff University, sorts data acquisition into three families: contact, noncontact and destructive.

Contact methods: touch probes and CMMs

A coordinate measuring machine (CMM) carries a probe on three perpendicular axes, each with a built-in reference scale. As the probe touches the surface, the machine records coordinates. There are two main styles:

• Touch-trigger probes record one point per touch. They're mounted on a CMM or on a manually guided articulated arm. The arm has more freedom of movement for complex parts, while the CMM is more accurate.
• Analogue scanning probes stay in contact and slide along the surface, recording continuously. The book says this is up to three times faster than point-to-point sensing.

Contact methods are accurate, relatively cheap, unaffected by color or transparency, and able to reach into deep slots and pockets. Their weaknesses are speed, since points are captured one at a time, and contact pressure, which can deform soft materials such as rubber. Chapter 1 puts contact scanning accuracy at roughly 0.01 to 0.02 mm. The Tennessee chapter notes that ultra-high-end CMMs reach 1–2 microns, while more affordable ones work in the 10–20 micron range.

Noncontact methods: light, sound and X-rays

Noncontact systems project energy, usually light, onto the object and measure what comes back. They capture data far faster and never touch the part. In exchange, they typically lose some accuracy (Chapter 1 cites about ±0.025 to 0.2 mm), and they struggle with shiny, transparent or dark surfaces. That's why technicians sometimes dust a part with a temporary powder coating before scanning.

The main noncontact techniques are worth understanding one at a time.

1. Laser triangulation

A light source projects a spot or line onto the object at a known angle, and a camera offset by a known baseline distance sees where the light lands. Because you know the triangle's base and angles, you can calculate depth. Chapter 3 gives the relationships, where F is the camera's focal length, L the baseline, θ the projection angle and P the image coordinate of the illuminated point:

Z = (F · L) / (P + F · tan θ)
X = L − Z · tan θ

Error in depth:
ΔZ ≈ (Z² / (F · L)) · ΔP  +  (Z² · sec² θ / L) · Δθ

The error equation is the one to remember. Depth error grows with the square of distance and shrinks as the focal length and baseline increase. Double the standoff distance and your error roughly quadruples. That's why triangulation scanners suit small objects at short range, and why the baseline can't simply be made huge: the scanner's physical housing limits it.

The Tennessee chapter adds a practical rule: the camera-to-laser viewing angle usually sits between 30° and 60°. Near 0°, the camera can't see the laser on flat surfaces. Near 90°, it's looking straight down the laser plane and can't detect depth changes. The middle range balances visibility against depth sensitivity.

2. Structured light

Instead of one spot, the system projects a pattern: a single line (a "sheet of light"), multiple stripes, a grid or a coded pattern. The camera captures how the pattern bends over the surface. A sheet of light captures hundreds of points along a line at once, which is why it's so much faster than a single scanning spot.

The Tennessee chapter compares the three illumination types for an image of M × N pixels:

• A single-point laser needs M × N measurements, one for each pixel.
• A sheet of light needs only M measurements, one per line.
• A coded pattern needs roughly log(M) patterns for similar resolution.

Coded light looks fastest on paper, but switching patterns becomes the bottleneck in practice, so sheet-of-light systems were the speed leaders when the book was written. Chapter 3 notes that structured-light systems became the favorite for scanning people because they're very fast, capture color texture, and don't need a laser.

3. Time of flight and phase shift

A time-of-flight (TOF) scanner fires a laser pulse and times its return. Because light travels at a known speed C, distance follows directly:

D = (C × t) / 2

You divide by two because the light makes a round trip. TOF is the tool for big, distant objects such as buildings and bridges, with accuracy from a few millimeters to a few centimeters at long range. A variant measures the phase shift between emitted and returning light instead of a pulse's travel time, which gives higher accuracy. Because the light travels out and back along the same line of sight, phase-shift systems can also see into holes and cavities.

4. Other techniques

• Moiré interferometry analyzes interference fringes. It's accurate on small surfaces but trades precision for range on larger ones.
• Passive methods, such as shape from shading and stereo vision, use only ordinary images. The hardware is simple, but finding matching points between images (the correspondence problem) is hard, and accuracy is limited.
• Coherent laser radar measures large-scale geometry precisely. The Metris system cited in the book reaches 16 µm at 1 m.
• Sonar and radar are inexpensive and long-range but low in accuracy. They're used for mapping, not precision parts.
• CT and MRI see inside objects. Industrial micro-CT can resolve details of tens of microns, even in dense materials. These are the most expensive options in both hardware and software.

Destructive methods: slicing the part

The most counterintuitive technique destroys the object. The part is embedded in a contrasting colored resin, a CNC mill shaves away one thin layer at a time, and a camera photographs each newly exposed cross-section. Software stacks the slices into a 3-D model.

Chapter 3 notes this is the mirror image of 3-D printing: additive manufacturing builds a part layer by layer, and destructive scanning takes one apart layer by layer. The cited system had layer thicknesses from 0.0127 to 0.254 mm and captured internal features that no camera could reach. The obvious cost is that you lose the part.

Contact vs noncontact vs destructive: a quick comparison

How do you turn a point cloud into a CAD model?

Collecting points is only the start. The Tennessee chapter describes a processing pipeline, and each step has its own math.

Step 1: Calibration

A scanner's raw output is a range image, where each pixel (i, j) has a range value r in arbitrary units. Calibration finds a transform T that converts it into real-world coordinates:

p = (x, y, z) = T(r, i, j)

Without calibration, you have a picture of depth. With it, you have measurements in millimeters.

Step 2: Registration

No single view can see every surface. A sphere hides its back from any one viewpoint, a problem called self-occlusion. Triangulation scanners have it twice over, because both the laser and the camera need a clear line of sight. So you scan from several angles, and each scan comes back in its own coordinate system. Registration finds the rigid transformation that aligns them:

p = R · p′ + t

Here R is a rotation matrix and t a translation vector. The scans must overlap. With no shared region, the algorithm has nothing to align. In practice, Chapter 3 describes a two-stage approach: a person first picks matching landmark points (three is enough to fix the orientation) for a rough alignment, and an automatic algorithm then fine-tunes it within a tolerance.

Step 3: Cleaning and optimizing the points

Every measurement carries error. The Tennessee chapter writes it as:

q = p + e

where p is the true point, q the measured one and e the error. Processing software reduces that error and trims the data set:

• Noise reduction moves stray points toward where they statistically belong. Settings differ for organic shapes, which should stay smooth, and mechanical ones, which need to keep sharp edges.
• Redundancy reduction removes points that overlap after merging scans.
• Sampling cuts the point count in one of three ways. Curvature sampling keeps more points in tightly curved regions. Random sampling removes a percentage of points. Uniform sampling divides space into equal cubes and keeps one point per cube.
• Primitive detection recognizes planes, cylinders and spheres and replaces noisy points with mathematically perfect shapes. This is especially useful for machined parts.

Step 4: Mesh reconstruction

Next, you "connect the dots" into a mesh of triangles. That sounds trivial, but it's a hard research problem for scattered points. Triangles became the standard because they're always convex and easy to render.

The Tennessee authors name three factors any reconstruction algorithm has to handle: the object's topology (a doughnut has a hole that a ball doesn't), the overlap between scans, and missing data, which leaves gaps in the mesh. Sampling density is a trade-off. Too few points gives an inaccurate model, and too many overwhelms your computer.

Mesh cleanup tools then fill holes, sharpen edges the scanner rounded off, refine regions that need more triangles, and decimate regions that need fewer. They also fix bad geometry, such as non-manifold faces (an edge shared by more than two faces), redundant faces and crossing edges. A model needs to be "watertight" before you can 3-D print it.

Step 5: Surface fitting with NURBS

A triangle mesh is fine for 3-D printing, animation and many simulations. Engineering CAD, however, usually needs smooth mathematical surfaces, typically NURBS (non-uniform rational B-splines). A parametric surface maps two parameters onto 3-D space:

S(u, v) = ( f(u, v), g(u, v), h(u, v) )

Chapter 3 explains why NURBS dominate CAD. They represent both simple analytic shapes and free-form shapes in one mathematical form, they store complex shapes compactly, they evaluate quickly and stably, and they stay consistent under the transformations CAD software applies. They also generalize B-splines and Bézier curves, so other common curve types fit inside the same framework.

Complex objects can't be described by a single surface, so the model is split into patches. Each patch gets its own grid of control points, and patches are trimmed to fit irregular boundaries. The book describes three ways to build NURBS models:

• From basic CAD entities. Extract points, curves and primitives from the scan, then model the part in CAD with them as references. This suits mechanical parts.
• From manually laid-out patches. Trace a layout of four-sided patches over the mesh. The book calls this the most effective approach for very complex shapes such as anatomy.
• Automatically from the mesh. This is a one-click conversion. It's quick, but the book cautions it suits only simple shapes or rough drafts.

The number of control points is a trade-off too. Too few gives an inaccurate surface, and too many gives a huge file. You check your result with an error map, a color-coded plot of how far the fitted surface deviates from the original scan.

The four software phases

Chapter 3 summarizes the software chain in four phases, which is a handy mental model for students:

1. Points and images: register, clean and optimize the raw data. For CT and MRI, this includes segmentation, separating the object of interest by gray-level thresholds and region growing.
2. Polygon: build and refine the triangle mesh.
3. Curves: extract cross-sections, feature lines and primitives.
4. NURBS surfaces: produce the final CAD-ready model.

Where you stop depends on the job. A rapid prototype needs only a clean mesh in STL format. A part you plan to redesign needs NURBS. Common output formats include IGES, STL, DXF, OBJ, VRML and CNC G-code.

Why is a scanner's accuracy number misleading?

This section of Chapter 4 is the one I'd make every student read twice. When a vendor quotes a single accuracy figure, it hides a lot:

• Accuracy varies across the working volume. Triangulation gets worse with distance, as the error equation showed. A robot arm flexes more at full reach.
• Mean vs RMS. Some vendors quote the mean error and others the root-mean-square. The mean is always the smaller, more flattering number, so ask which one you're seeing, and which test protocol produced it.
• Resolution limits accuracy at edges. Suppose a scanner records one point per millimeter. If one point lands just outside an edge, the next lands almost a millimeter inside it. Even if each point is measured perfectly, your edge position could be off by up to 1 mm.
• Errors stack up. If a part is too big for one scan and each registration adds 0.1 mm of error, joining ten sections could put you a full millimeter out.
• Lab conditions differ from your shop. Vendors test in controlled environments. Temperature, humidity and operator technique all change real-world results.

The book's answer is to test the whole chain. Scan a part, build the CAD model, manufacture a replacement, then measure both the original and the copy (with a CMM, for example) and compare them statistically. The real question isn't "how accurate is the scanner?" It's whether the final output is fit for its purpose. And no amount of clever processing can add accuracy that wasn't captured in the first place.

How do you choose a reverse engineering system?

Chapter 4 also lays out a structured buying process, which is useful even if you never buy a scanner, because it's a solid template for any major technology decision:

1. Form a team. A senior project champion, a small full-time "inner team" and a wider "outer team" of stakeholders who stay consistent throughout the process.
2. Write the business case. Define the opportunity, technical requirements, budget, and whether the volume of work justifies an in-house system or a service bureau.
3. Define requirements and how you'll test them. Rank and weight attributes. For example, decide in advance whether speed matters more than accuracy.
4. Gather vendor information. The authors note, amusingly, that searching for "reverse engineering" alone mostly returns software decompilation tools. You need terms such as "scanner" or "digitizer."
5. Short-list using quick yes/no questions, including whether the scanner, point-cloud software and CAD tools actually work together.
6. Visit vendors, assess them in detail, and run benchmarks using your own parts in conditions close to your real environment.
7. Evaluate the vendor commercially: financial stability, reputation and local support.

How are reverse engineering and 3-D printing connected?

Reverse engineering and rapid prototyping (RP), what we'd now usually call 3-D printing or additive manufacturing, are natural partners. RE goes from physical to digital, and RP goes from digital back to physical.

Chapter 5 describes the RP process as six steps: create a CAD model, convert it to STL (a mesh of triangles), slice it into thin layers, add support structures for overhangs, build the part layer by layer, and post-process it. Techniques covered include stereolithography, selective laser sintering, fused deposition modeling, 3-D printing, laminated object manufacturing, multijet modeling and laser-engineered net shaping.

Chapter 6 tackles a subtle problem. If you go from point cloud to NURBS to STL to slices, errors pile up at three separate stages:

1. between the point cloud and the surface model,
2. between the surface model and the STL mesh, and
3. between the STL mesh and the printed layers.

Because the stages are separate, you can't control the total error well. The chapter presents an approach from a team at the National University of Singapore that skips the intermediate models. It slices the point cloud itself into layers, fits a 2-D contour to each layer's points, and adjusts each layer's thickness so the shape error stays just within a tolerance you set. That gives one error to manage instead of three, and close to the minimum build time. It's a nice example of a general engineering lesson: every format conversion is a chance to lose information.

What do real reverse engineering projects look like?

Chapters 7 to 9 are case studies. Keep in mind that they describe the state of things when the book was written, so the figures are historical, not current.

Automotive

• Faster car body design. The book says a full-scale car design traditionally took more than three months and that Japanese automakers were working to cut that to three days. The workflow starts with a quarter-scale clay model, scans it, and builds the surfaces digitally.
• GM's NASCAR engine block. Richard Childress Racing needed a digital model of GM's SB2 racing engine block, whose old 2-D drawings didn't reflect changes made to the casting molds over the years. As the project engineer put it, they wanted a model of what they actually had, not an idealization of it. To capture internal water-jacket cavities that no line-of-sight scanner could see, the team poured urethane into the actual sand-core molds and scanned the castings. The final model was three times more accurate than the casting tolerance required: ±0.010 inch versus ±0.030.
• Ferrari at Le Mans. To redesign the rear wing of the Veloqx Prodrive Ferrari 550 Maranello, the team scanned one half of the car body (about 40 million points), mirrored it to create the other half, built NURBS surfaces, and ran airflow simulations. A new wing was ready for track testing in six weeks and cut drag by 2.5% at the same downforce.

Aerospace

• Legacy parts. One service provider quoted in the book estimates a company is lucky if it has digital models of even 1% of its inventory. The book also cites the economic case for keeping old fleets flying: less than 10% of the investment in a new airplane can keep 80% of aging airplanes in service.
• NASA's X-38. The prototype was roughly 33 feet long, far too big for a CMM, and it had to be scanned in a working vehicle bay with changing light, moving doors and only three feet of clearance underneath. No anti-glare spray was allowed. The team combined photogrammetry markers with a white-light scanner. Measurement and processing took four days, and a CAD model for airflow analysis took another eight hours.
• Faster inspection. Supplier AMT cut first-article inspection setup from about half a day to as little as half an hour by switching from CMMs to laser scanning and automated comparison against CAD, and started building a library of 200 reusable inspection templates.

Medical

• Invisalign. Align Technology's clear aligners depend on scanning dental impressions. Its first attempt used destructive slicing: plaster models encased in black material, hardened for about 17 hours, then shaved layer by layer. It worked for research but was too slow and wasteful for production. The company built its own structured-light scanners, reaching 0.1 mm accuracy, and later moved to desktop industrial CT, which scans the impression directly and handles several at once.
• Digital dentistry. KaVo's Everest system scanned a tooth stump in about four minutes with 20-micron accuracy, combining 15 views automatically.
• Hearing aids. GN ReSound scanned ear impressions to produce custom shells and reported a 75% reduction in labor.
• An artificial heart. Cleveland Clinic engineers scanned a plastic model of a blood pump chamber with a white-light scanner. They had tried a CMM first, but it was too slow and not accurate enough.

The Invisalign story is my favorite teaching example because it shows the whole technology trade-off space in one company: destructive versus noncontact versus CT, accuracy versus throughput, and buying versus building your own tools.

Is reverse engineering legal?

Chapter 10 covers legal aspects. It focuses on U.S. software copyright law, on the theory that software cases would set the tone for reverse engineering in other fields. This is not legal advice, and the law has changed since 2008, but the core cases are still taught today.

The key ideas

• Reverse engineering is legitimate in principle. The U.S. Supreme Court, in Kewanee Oil Co. v. Bicron Corp. (1974), described it as a fair and honest means of starting with a known product and working backward to understand how it was developed or made.
• Copyright protects expression, not ideas. U.S. copyright law doesn't cover ideas, procedures, processes, systems or methods of operation. Function can be protected by patents, not copyright.
• Sega v. Accolade (1992). Accolade disassembled Sega cartridges to make compatible games without a license. The Ninth Circuit held that when disassembly is the only way to access a program's ideas and functional elements, and there's a legitimate reason for that access, disassembly is fair use.
• Atari v. Nintendo (1992). The Federal Circuit agreed in principle that you can study a work you rightfully own to understand how it operates. But Atari had obtained Nintendo's source code by misrepresenting itself to the Copyright Office, and that "unclean hands" sank its fair use defense.
• Lotus v. Paperback (1990). A court found that Paperback's spreadsheet infringed by copying too much of Lotus 1-2-3's look and feel rather than just its functional elements.

The practical rules the book draws

1. Start from an authorized copy bought on the open market, and honor any license agreements you've signed.
2. Use reverse engineering only when there's no other way to get the information, for example when the vendor doesn't publish an interface.
3. Take only what you need, meaning the functional elements, not the creative expression.

What has changed since 2008

Two developments matter for students today. In the U.S., the Digital Millennium Copyright Act of 1998 restricts circumventing technical protection measures, which adds a layer beyond the fair use analysis. In the EU, the Software Directive (2009/24/EC) specifically permits decompilation for interoperability under strict conditions. Physical products bring patents, design rights, trade secrets and contracts into play too. Before any real project, check the rules where you are and the license of what you're analyzing.

Why don't more companies adopt reverse engineering?

The final chapter studies what drives adoption among UK small and medium manufacturers, who make up the vast majority of the sector. Using semi-structured interviews with RE experts, the study tested factors across organizational, environmental and project dimensions.

The factors that mattered were a champion inside the company, top management support, genuine internal need, the degree of competitive pressure, coordination of resources, and end-user participation. Two factors that you might expect to matter, company size and help from external consultants, didn't show a significant effect.

Read those results with caution. About 80% of the sample hadn't yet adopted the technology, so effects that appear only during implementation may have been missed. Still, it's a useful reminder that technology adoption is usually an organizational problem before it's a technical one.

How does this connect to software reverse engineering today?

Here's where my software background comes back in. The mechanical pipeline and the software pipeline are surprisingly similar:

Both disciplines raise the abstraction level step by step, and both lose information at every conversion. Both are now being automated. Laser scanners automated what CMM operators did by hand, and AI agents are now automating what reverse engineers do in Ghidra or IDA. I covered that in my guide to REA, a reverse engineering tool for AI coding agents. The same warning applies to both: automation speeds up the work, but you still need to verify the output and stay within the law.

Glossary for students

• Point cloud: a set of 3-D coordinates sampled from an object's surface.
• CMM: coordinate measuring machine, a precise touch-probe measuring system.
• Triangulation: calculating depth from a known baseline and angles between a light source and a camera.
• Structured light: projecting known patterns and analyzing how they deform over a surface.
• Time of flight (TOF): measuring distance by timing a light pulse's round trip.
• Occlusion: part of an object hiding another part from the scanner's view.
• Registration: aligning multiple scans into one coordinate system.
• Mesh: a surface made of connected triangles.
• Watertight: a mesh with no holes, required for 3-D printing.
• NURBS: smooth parametric surfaces used in CAD.
• STL: the standard triangle-mesh file format for 3-D printing.
• Error map: a color-coded plot of deviation between two models.
• Fair use: a U.S. legal defense permitting limited use of copyrighted material.

Questions to test your understanding

1. Using the triangulation error equation, explain what happens to depth error when you double the scanner's standoff distance. What two design parameters reduce it?
2. Why does registration require overlapping scans? What would happen with no overlap?
3. A scanner captures one point per 0.5 mm with near-perfect point accuracy. What's the worst-case error in locating a sharp edge?
4. You need to digitize a soft rubber gasket with an internal channel. Which acquisition methods would you rule out, and why?
5. Name the three stages where error accumulates when converting a point cloud into a 3-D printed part, and describe how adaptive slicing of the point cloud reduces this.
6. Why did Atari lose its fair use defense even though the court accepted that studying a program can be fair use?
7. Choose one case study and identify which scanning technology it used, why it was chosen, and what the alternative would have cost.

Reverse engineering looks like a shortcut from the outside: scan it, copy it, done. Up close, it's one of the most demanding parts of engineering, combining optics, geometry, statistics, software and law. If you understand why depth error grows with the square of distance, why a vendor's accuracy number needs context, and where the legal lines sit, you're already ahead of most people who use these tools.

Sources

• Raja, V. and Fernandes, K. J. (Eds.) (2008). Reverse Engineering: An Industrial Perspective. Springer Series in Advanced Manufacturing. Springer-Verlag London. ISBN 978-1-84628-855-5.
• Chapters cited: Page, Koschan and Abidi (Ch. 2, methodologies and 3-D laser scanners); Pham and Hieu (Ch. 3, hardware and software); Ch. 4 (system selection); Ch. 5–6 (rapid prototyping and adaptive slicing); Ch. 7–9 (automotive, aerospace and medical case studies); Ch. 10 (legal aspects); Ch. 11 (barriers to adoption).
• Kewanee Oil Co. v. Bicron Corp., 416 U.S. 470 (1974); Sega Enterprises Ltd. v. Accolade, Inc., 977 F.2d 1510 (9th Cir. 1992); Atari Games Corp. v. Nintendo of America, Inc., 975 F.2d 832 (Fed. Cir. 1992).