Inferring 3D structures from 2D photographs is considered an “ill-posed” problem in computer vision in the sense that its solution is inherently under-constrained and unstable. Visualsize's Photomodel3D employs an analysis principle called “structure from motion” for such a 3D inference process, and the process comprises many intertwined components. Failure of any of the components can defeat the whole modeling procedure. Hence, to ensure success, some caution should be exercised in gathering pictures for a 3D analysis.

What Types of 3D objects Can Be Modelled?

Please take a look at our demo pages and interactive demo pages. While our Photomodel3D program works with many different types of objects, the important thing to remember is that it does not work well with objects whose appearance is too plain (a single, or a few, colors with no texture), shinning (glass and mirror), or changing (moving or deformable objects). Also, data used for modeling must be real images and videos from a camera, a phone, or a camcorder; not animation sequences synthesized by some computer graphics programs.

What You Get

We provide 3D models in two different resolutions and four different formats:

(1) A high-resolution model in wrl format (mesh.wrl) that can be viewed directly inside a Web browser. You need to have a proper browser plugin installed to view wrl files. Please visit National Institute of Standards and Technology to select and perform a one-time installation of a browser plugin.

(2) Alternatively, you can download the high-res model in compressed ply format (meshCOLOR.ply.gz, compressed to save bandwidth) and use a CAD program to view and edit the model on your own computer. For non-commercial use, we recommend Meshlab

(3) A low-resolution model in WebGL format (index.htm) that can be viewed directly inside a web browser that supports WebGL. Note that you must have a graphic card with the right driver to view a WebGL model. Furthermore, many of our low-res models still have too many vertices to be displayed using WebGL. So this low-res model may or may not be available for your data sets.

(4) A low-resolution model in compressed ply format (mesh.ply.gz) for downloading and viewing on your computer.

(5) A model in compressed stl format (mesh.stl.gz) for use with most single-color 3D printers.

Please visit our interactive demo page to view sample high-res and low-res 3D models.

Basic DOs and DON’Ts

A project needs at least three images. We recommend at least 5 to 10.

Use an image resolution about 1M to 3M pixels (e.g., 1280x960, 1600x1200, and 2048x1536) to provide sufficient surface details. Images smaller than VGA (640x480) will be rejected by the GUI. Images of resolutions higher than 3M pixels can take much longer to upload and process but do not necessarily provide more details.

Make sure that there is sufficient environmental lighting when taking pictures (outdoors are generally better than indoors). However, do NOT use flash. If you need flash to take good pictures, then the environment is too dark for a good 3D reconstruction.

Just point and shoot, do not change any camera parameters (zoom, aperture, etc.).

Make sure that adjacent pictures show significant overlap and the direction where the camera points at does not change significantly from one photo to the next. The rule of thumb is that the part of the 3D scene that you would like to model should appear in at least 3 adjacent photographs.

Maintain a smooth camera trajectory in taking picture, not one random picture here and another there. Generally speaking, it is difficult to guarantee a good, complete spatial coverage if pictures were taken at random locations around an object.

The object of interest should occupy as large a portion of the photograph as possible. (If you want to model a certain 3D object, pixels not depicting that object are wasted.) For example, to generate a 3D model of a figurine, your pictures should look like the one on the right, not the one on the left where the object is way too small, of the following illustration. You can use either portray or landscape modes.

Do not stand at a single location to take pictures. Walk around the object of interest and take pictures from all around the objects. However, NEVER put the object of interest on a turntable and rotate the object to reveal different faces to a stationary camera.

You will need about 30 photos for a 360^o all-around model construction. Take 50 to 60 photos can better ensure success. Furthermore, for 360^o all-around model construction, it is much better to circle the object ONCE instead of multiple times. For example, if you are to take 60 pictures all around, instead of circling the object three times and take 20 pictures each time, you want to circle the object just once to take all 60 pictures. The reason is that the change of the viewing direction in between adjacent pictures will be much smaller and predictable in the latter.

Additional Instruction For 3D Face Models

Change the project type to "dense model".

Use the portray instead of landscape mode, and make sure that the face is as big as possible. For example, here are some examples of good images (big head shot in portray mode), and here are some examples of bad images (small head shot in landscape mode).

Collect about 15 to 20 images by taking pictures from left to right (or right to left), roughly orbiting the front of the subject. The subject should not move, giggle, laugh, or talk when pictures are taken. Here is an example of a good data set.

VisualSize offers the best-in-class technologies for reconstructing 3D information from regular digital pictures. Our technologies use pictures from mass-market, consumer-grade digital cameras, camcorders, and phones without the need of special equipment, elaborate set-up, advanced calibration, man-made markers, artificial lighting, user training, or manual intervention. To request advanced and customized versions, please contact us at To learn more about Visualsize's 3D offerings and how Visualsize compares to the competitors, please peruse our technology overview slides.