From Photos to 3D Models! SfM Processing Steps and Tips to Avoid Failure

In recent years, the technology to create 3D models from photos has advanced dramatically. The representative method, SfM processing (photogrammetry using Structure from Motion), is attracting attention because it can build high-accuracy 3D models from photos taken with ordinary cameras or drones without special equipment. The introduction of photogrammetry is progressing in construction and surveying sites for situational awareness and earthwork volume calculations, and environments where even beginners can easily try it are becoming available.

This article clearly explains everything from the basics of SfM processing to photo shooting tips, software processing flow, and tips for improving accuracy. It includes key points to help those trying SfM processing for the first time or engineers who want to use photogrammetry in their work avoid failure. At the end of the article, we also introduce how to support SfM processing using LRTK, a simple surveying tool.

Basics of SfM Processing: How Photogrammetry Creates 3D Models

SfM (Structure from Motion) processing is a photogrammetric method that reconstructs the three-dimensional structure of a subject from multiple photos. Traditionally, 3D measurement required expensive devices like laser scanners (LiDAR), but with SfM processing you can generate 3D models using only photos taken with commercially available digital cameras, smartphones, or drones. Thanks to advances in image analysis algorithms, common points between images can be identified to estimate camera positions and orientations, and the object’s shape can be calculated by triangulation.

To briefly explain the mechanism of photogrammetry, the same feature points (distinctive markers) that appear in multiple photos are used to determine each photo’s camera position and orientation at the time of shooting. For example, if a feature point appears in two or more photos, direction rays from each camera position to that point can be drawn and their intersection determined by triangulation to compute the 3D coordinates of the feature point. Performing such calculations for countless points yields a 3D point cloud (a collection of feature points) that forms the skeleton of the model. By densifying the point cloud using information from multiple corresponding photos and converting it into a polygon mesh, you can construct a 3D model that reproduces continuous surface shapes and textures of the object.

The advantages of SfM processing include that it does not require specialized equipment and is low-cost and easy to start. As long as you can shoot photos and perform image processing on a PC, anyone can generate 3D data. This has expanded use across various fields such as terrain surveying at construction sites, documentation of civil engineering structures, cultural heritage archiving, equipment inspections, and creating models for CG. Also, point cloud data from SfM processing includes color information, making it attractive for visually based analyses such as concrete crack detection or vegetation analysis.

However, it should be noted that models obtained by SfM processing can reproduce relative shapes (proportions) but cannot determine absolute scale (dimensions) by this process alone. Since the actual size of the subject is unknown from photos alone, additional references are necessary to match actual dimensions or coordinate systems. For example, from photos of a model alone you cannot tell if it is the same size as the real object, but placing a tape measure or scale bar (a rod of known length) beside it provides a dimensional reference. Similarly, when using SfM for surveying, it is important to place known-distance markings or survey targets on site and photograph them together so the model can be scaled and aligned in subsequent steps.

Photo Shooting Tips Suitable for SfM Processing

It is not an exaggeration to say that the success of SfM processing is determined by how photos are taken. To prevent common beginner mistakes, keep the following points in mind during shooting. If each photo contains sufficient information, image analysis on the software side proceeds smoothly and a high-quality 3D model can be obtained.

• Shoot facing the subject – Point the camera as much as possible directly at the surface of the subject. Avoid extreme oblique angles and aim for angles where each part is clearly captured. Flat subjects like walls or ground can distort more when shot from oblique directions, reducing the accuracy of feature point detection. Shooting head-on lets you effectively use the full resolution of the photo and record fine details clearly.

• Ensure sufficient parallax – Take multiple shots while changing the camera position to secure large viewpoint differences (parallax) between photos. Rather than standing in one place and only rotating the camera, take a step and move around the subject while shooting. Greater parallax stabilizes depth estimation by triangulation and improves accuracy. However, if you move too much and adjacent photos share few common areas, problems arise, so balancing parallax while maintaining the overlap described below is important.

• Take ample overlap between photos – Consecutive photos should be taken with significant overlap. Ideally, adjacent photos should share 60–80% or more common area. The more overlapping regions captured, the easier it is for the software to find matching points and align the photos correctly. When shooting wide areas, divide the area systematically and shoot with a high overlap rate. For example, for drone aerial photography, forward overlap of 80% and side overlap of 60% are recommended.

• Do not change the light-source–subject relationship during shooting – Be careful to avoid extreme changes in lighting conditions during a shooting session. Outdoors, finish shooting within a time window where the sun’s position does not change significantly; indoors, keep artificial lighting constant so that shadow directions and brightness do not vary drastically across the photo set. Large brightness differences between photos make it harder for the software to detect matching feature points and can leave strong shadows in the model that complicate post-processing. Ideally, photograph the entire subject under consistent lighting conditions.

• Avoid blur and camera shake – Taking sharp, well-focused photos is fundamental. If out-of-focus or blurred images are mixed in, those images may yield no feature points, creating holes in the model or large alignment errors that degrade overall accuracy. Use a tripod, shoot during bright hours, use short shutter speeds and appropriate aperture settings, and take advantage of camera stabilization features to collect the clearest possible photos. When shooting in sequence, lock camera settings (focus, exposure, etc.) to obtain consistent images and stable results.

Considering these points, photographing the subject thoroughly from all parts is the secret to successful SfM processing. For small objects, walk all the way around and shoot from various angles; for large structures or terrains where you cannot go around, take photos from multiple directions as much as possible. For example, when modeling a building’s exterior wall, move the camera parallel to the wall several times while changing vertical angle to capture each part well. If you capture enough photos to cover the entire subject, subsequent software processing will be much smoother.

Software Processing Flow for SfM

Once your photos are prepared, you can run SfM processing with software. There are many user-friendly photogrammetry software options today, and the basic workflow is common. Here is a step-by-step explanation of the typical processing flow.

• Import photos and detect feature points Load the multiple photos you took into dedicated SfM software. The software automatically detects feature points in each image (such as high-contrast points and corners) and matches common points between photos. This feature point matching information becomes the basis for subsequent 3D reconstruction. While more photos increase computation, the computer will automatically find matches across wide areas, so this is manageable.

• Photo alignment (camera pose estimation) Based on feature point correspondences, the software estimates each photo’s camera position (coordinates) and orientation (pose) at the time of shooting. The software incrementally solves the camera layout from no initial values and ultimately determines the relative positions and directions of all photos. Simultaneously, each feature point’s spatial location is computed by triangulation, generating a sparse point cloud (a low-density 3D point cloud). This is the core part called SfM processing; after processing, photos are spatially placed relative to one another, and camera markers and the point cloud are displayed on the screen to recreate the shooting scene.

• Point cloud densification (MVS processing) After photo alignment is complete, proceed to densify the point cloud. This step, called MVS (Multi-View Stereo), analyzes correspondences across multiple photos for each pixel to obtain finer 3D points. MVS processing transforms the sparse point cloud into a dense point cloud (which can reach millions to hundreds of millions of points), allowing detailed representation of the subject’s shape. For example, fine architectural decorations or subtle terrain undulations can be captured by this dense point cloud. The process takes some time but is essential for a detailed model.

• 3D mesh modeling and texture generation Once you have a dense point cloud, generate a polygon mesh model as needed. Point clouds are collections of discrete points; by connecting them into surfaces (meshes), you produce a continuous 3D model. Meshing represents the shape with smooth polygons, and projecting original photos to create textures produces realistic 3D models that look like the photos. This mesh + texture can be displayed in 3D viewers or exported in formats usable by other CAD software (e.g., OBJ format, point clouds as PLY/LAZ formats). For surveying purposes texture may not be necessary, but textured models are useful for site records such as crack detection.

• Scale adjustment and coordinate alignment (if necessary) Models from SfM processing are initially at an unknown scale. To create a full-scale 3D model, perform scale adjustment (scaling). Specifically, specify a known distance measured on site (for example the distance between markers placed on site or the length of a ruler included in photos) within the software to match the model’s dimensions to that real distance. If you want to work in a surveying coordinate system, you can also assign control point coordinates to points on the model for georeferencing (coordinate alignment). Many photogrammetry software packages provide functions for scale setting and coordinate transforms; by inputting values, you can align the model to real-world dimensions and coordinate systems. For example, if you set two points to be “5.000 m (16.404 ft) apart in reality,” the software applies that scale to the entire point cloud, enabling measurement in actual units thereafter.

• Output and use of deliverables After processing and scaling, output the necessary deliverables. Typical results from SfM processing include 3D point cloud data, mesh models, and orthophotos (orthorectified images) viewed from above. In civil engineering and construction, cross-sections and height maps can be created from point clouds or orthophotos, and earthwork volume calculations can be performed. In cultural heritage fields, mesh models can be archived and used for VR/AR exhibits. For on-site quick measurements, use measurement functions on point clouds or orthophotos to obtain distances. Because the 3D models from SfM processing are a “digital copy” of the site, they can be used for diverse analysis, documentation, and design review.

Tips to Improve the Accuracy of SfM Processing

Anyone can generate a 3D model by following the basic steps, but when high accuracy is required for surveying purposes, additional efforts are needed. Here are points and precautions to improve the accuracy of SfM results.

• Prepare enough photos and high-quality images: Model accuracy depends on the amount of information obtained from photos. Shoot the subject from various angles and secure a sufficient number of photos. Ensure the same area appears in at least 2–3 photos from different angles; for complex shapes this can mean hundreds of photos. Use the highest-resolution camera possible and maintain image quality (resolution, sharpness, proper exposure). Low-quality images reduce point cloud fidelity and affect dimensional accuracy and surface detail.

• Countermeasures for low-feature areas: Surfaces that do not provide feature points—glass, mirrors, water surfaces, or plain white walls—cause holes or distortions in SfM processing. Add artificial patterns such as sticky notes or markers before shooting, or change shooting angles to avoid reflections. Also, avoid moving objects (leaves blowing in the wind, cars or people) as much as possible; shoot in a static environment. Removing unnecessary objects beforehand is another option. Creating an environment where sufficient feature points are obtainable leads to improved accuracy.

• Camera calibration: SfM software generally estimates internal camera parameters (focal length, lens distortion coefficients, etc.) simultaneously during processing. However, for higher accuracy you can perform pre-calibration to correct lens distortion in advance. Commercial photogrammetry software and open-source libraries allow you to calibrate by photographing a checkerboard pattern and applying those results to processing. This is not mandatory, but it is recommended when using wide-angle lenses or aiming for high-precision measurements.

• Use of control points and known coordinates: Related to scale adjustment, using reference points called Ground Control Points (GCPs) greatly stabilizes accuracy. GCPs are points on site with accurately measured coordinates (geodetic latitude/longitude or planar coordinates), and associating these with photos corrects the model to absolute accuracy. For example, place conspicuous markers on the ground and survey their positions with GNSS, then assign those coordinates to corresponding points in the SfM software. With multiple GCPs, the model’s position, rotation, and scale will match reality, yielding accuracy suitable for surveying. In sites with significant elevation changes, distribute GCPs across the area to reduce three-dimensional distortions.

• Verify model quality: After processing, always check result quality. Some software reports numerical measures such as reprojection error (projection mismatch on photos) and GCP residuals. If large errors are reported, some photos may be misaligned; remove those photos and reprocess or take additional photos to fill gaps. Also, measure known distances on the resulting point cloud or model to confirm they match expectations. If necessary, re-adjust scale or add more GCPs to correct the model until it meets the required accuracy.

• Post-processing according to measurement objectives: Point clouds and models generated by photogrammetry may contain noise points or missing parts. As part of improving accuracy, consider noise removal and hole-filling. Be cautious with excessive smoothing after meshing because it can distort dimensions; for measurement purposes, working with the point cloud is generally safer. Conversely, for presentation-focused use, retopologize the mesh to reduce polygon count and edit as needed. Apply processing appropriate to your objectives to achieve practical-quality 3D data.

By following these tips, the accuracy of models created by SfM processing will improve dramatically. It may take trial and error at first, but gaining experience from planning shoots, data processing, and verification will help you acquire the necessary skills. Preparation and checks are vital to “avoid failure.” Plan and shoot carefully and perform appropriate data checks to obtain high-accuracy 3D models.

Using Simple Surveying and LRTK to Support SfM

Finally, we introduce simple surveying tools to make SfM processing more convenient and accurate on site. Photogrammetry alone can yield high relative accuracy, but combining some surveying procedures is ideal for matching real-world coordinates or elevations and increasing reliability. Traditionally, control point surveying and RTK-GNSS positioning required specialized equipment and effort, but recently simple high-accuracy positioning systems combining smartphones and GNSS have appeared. One such solution is LRTK.

LRTK (smartphone RTK positioning system) is a simple surveying tool that enables centimeter-class position information acquisition. Using a dedicated smartphone app and a high-precision GNSS receiver, anyone can obtain high-accuracy positioning on site immediately. For example, using LRTK can support SfM in the following ways.

• Acquire photos with high-accuracy positions: When you take photos with an LRTK-compatible app, high-accuracy coordinates of the shooting position and camera orientation information are tagged to each photo. Importing this metadata into SfM processing speeds up and stabilizes photo alignment, and the model is automatically placed into geodetic coordinates, reducing the need for scaling. Even beginners can obtain geotagged images without complex settings, making SfM processing much easier.

• Simple on-site surveying (GCP acquisition): With LRTK, the traditional requirement for total stations or survey-grade GNSS to measure Ground Control Points is simplified. For example, place several markers (targets) in the target area and measure them with a smartphone + GNSS to obtain GCP coordinates. Then take photos as usual for SfM processing and assign those coordinates within the software to instantly relate the model to real space. This enables anyone to perform surveying, lowering the barrier to SfM adoption on site even without a dedicated surveyor.

• On-site quality checks: LRTK systems often allow uploading acquired point clouds and photos to the cloud for immediate simple analysis and checking. You can display the point cloud on a smartphone or tablet right after shooting to confirm that the entire area was captured, and check for gaps or errors. If there are omissions, you can perform additional shooting on site, and you can quickly measure distances or areas of interest from the point cloud. This on-site immediate feedback prevents discovering missing shots or insufficient accuracy only after returning to the office.

By combining simple surveying systems like LRTK with SfM processing, you can resolve SfM’s weaknesses such as unknown scale and accuracy concerns, enabling anyone to perform high-accuracy 3D measurement easily. For example, on construction sites, earthwork volume calculations that used to take a full day can be performed quickly using drone + SfM + LRTK, with results immediately shared via the cloud for stakeholder review. Even beginners can follow a smartphone app’s guidance to take photos and measure points, making it easy to perform 3D modelization during onsite work.

SfM processing from photos to 3D models is becoming an indispensable technology in the digitizing transformation of the construction and surveying industries. Use tools like LRTK that make adoption easier, and apply these techniques to advance on-site DX (digital transformation). With appropriate shooting and the latest technologies combined, you can achieve high-accuracy, fail-safe 3D model generation. Why not start digitizing your site with the camera at hand from tomorrow?