Recording images for 3D modelling
3D model of Port Royal National Historic Site (Nova Scotia, Canada) created from 600 ground and aerial drone images.
Soft and even lighting:
Try and have the object or scene evenly illuminated across the entire area of interest. If you are
recording the images outside under natural light, try and pick a bright overcast day, this will help eliminate dark
shadows and bright highlight areas. This even lighting will create a more pleasing and natural looking model.
Make sure you record all images under the same lighting conditions. If the lighting changes (sun comes out or
goes behind a cloud) different shadows will be crated on the object which will make it hard for the software to
determine the same points in different images.
Consistent exposure:
Try and keep the exposure the same for all images. If you are using a DSLR you might want to set the
exposure on Manual Mode and use an exposure setting that is average for all images. To use the same
exposure for all images you will need relatively even lighting across the entire area of interest. If you are using
a Point and Shoot camera you can use the exposure lock option and set the exposure for each image before
moving to different viewing angles. Try and watch for background objects that can cause errors in the
automatic exposure setting (white clouds, reflections, lights), these will cause problems if your camera is set on
auto exposure mode.
Small objects indoors:
If you are creating a model of a small object place it on a small stable platform that allows easy movement around the object (a card table in the middle of a well light room will work fine). Try and create even lighting around the entire object and don’t change the lighting during the image acquisition process. Place the object on a neutral coloured base and try and remove any objects close to the subject that might interfere with the modeling process. Record at least 30 images around the object and a couple rings of additional images above and below (if possible) the object. Placing the subject on a thin pedestal can help you record detail in the lower regions of the object. Make sure you don’t cast a shadow on the object as you move around the table.
Record lots of images:
Try and record as many images as you can of the same area from different angles. The quality of the
final model will be determined by the resolution of the images and the number of times a single part of the
subject is imaged from different angles. Every part of the model needs to be imaged from a minimum of 5
different angles so think through the process before you start recording images. If you are recording images
around a stationary simple objects (one without any strange indentations or protruding parts), make sure you
record at least 30 images around the object (more if possible). The more complex the object, the more images
you will need to record in order to avoid missing parts of the final model. Our brain fills in parts of the model for
us but the computer can’t do this so it needs images of all parts of the object in order to create an accurate
model. Try and record a ring of images at different heights to make sure you don’t miss spots, this redundant
collection of images will also help refine the final model.
Fill the frame:
Make sure you fill the frame with the object. Try to have the area of interest fill 90% of the frame. This
will allow the software to take full advantage of the resolution of the camera when creating the model. Use the
full resolution the camera is capable of recording and make sure the camera is set on the highest quality image
setting possible (Fine or Super Fine). If your camera allows for RAW image recording, it would be a good idea
to record this format as well as JPG images. Try and use as low an ISO setting as possible, this will give you
the best quality images.
Lots of overlap:
Make sure you have every part of your image recorded from many different angles (in 5 images or more if possible), this will ensure you don’t have
any parts of the image left empty due to lack of detail. These empty parts of the final model are called Data shaddows.
Clear and Sharp:
Make sure your images are clear and sharp. The software can only determine the locations of objects
it can clearly see, if an images is out of focus or blurred the software won’t be able to create an accurate
model. Make sure the subject doesn’t move during the image collection process. Move the camera,
not the subject.
Lots of texture:
The software is looking for unique points in the image that it can compare with other points in other
images so make sure your subject has visible texture. Areas like sand, snow, flat walls, water, windows will not
work well because the texture of the surface is to uniform or the surface is moving (in the case of water). Your
object needs to be completely still for all images in the series in order for the software to create a good model.
Objects like grass, water, trees, moving people or traffic will not create good models.
Patterns can cause problems:
Be cautious about imaging objects with reoccurring patterns. 3D modeling software is looking for unique textures and points so it may get confused if it sees lots of similar objects like identical bricks in a brick walls or reoccurring windows in a glass skyscraper.
Panoramas are not perfect:
Make sure you record images of the scene from different locations. If you stand in one position and
record a panorama, this series of images cannot be used for 3D model creation. BUT, if you stand at different
points and record panorama series from these points, these collections of panoramas can be used. In the case
of recording images inside a room, make sure you record panorama series from all 4 corners and additional
panorama series from the center point along each wall. Make sure you have at least 75% overlap in each
image for each panorama series. Youi may also need to record panorama series at different height in order to
fill in areas missed in the first series of images.
Turn OFF in-camera Image Stabilization:
The 3D modeling process relies on the fact that the lens and image sensor doesn’t move between
images (this is called the interior orientation) so if you have your in-camera image stabilization setting turned
on, the camera trues to shift the lens to compensate for any movement so this lens to sensor relationship is
constantly changing. This shift in lens to sensor causes errors in the final model and cause you problems when
you try and make accurate measurements afterwards. It is best to make sure you have lots of light so you can
use a higher shutter speed (less blur) or us a gimbal to stabilize the entire camera.
Record measurements:
Take a few seconds to measure the length of a few large objects in the scene. After you process the
images most software packages will allow you to assign a length to an object which will transform the model
into real world dimensions. Once this model has real world dimensions, you can now determine sizes of other
objects based on the reference length you entered.
Posted by Paul Illsley
|