Photography is an art. Photographing for 3D is an art de rigueur!
If we want to have a point in space to be reconstructed in 3D, this point has to be 'seen', or photographed, in two different viewpoints. This is exactly how we use our two eyes to see and perceive our 3D surroundings!
The reconstruction accuracy and quality is only limited by the image resolution, and by the camera distance to the object.
The higher the image resolution is, the higher resolution the reconstructed 3D model will be. It is recommended to fly lower, and take more images!.
Each image has to be overlapped with another image. This is equivalent to saying that the reconstructed areas have to appear in two photographs. It is recommended to have at least 70% overlapping between images. It will be the best to have 80%~85% overlapping. The 90% is good but it may be less efficient.
The overlapping images have to be at the different positions! It is recommended that you do not rotate the camera outwards while staying at the very close positions, which is good for 'panoramic' stitching, but is absolutely not for 3D reconstruction! The 3D reconstruction and orthomosaics are fundamentally different from the panoramic stitching or image mosaics.
Last but not least,
we can only reconstruct static subjects, not moving objects at all such as pedestrian and moving cars on the street.
The non-Lambertian materials such as transparent, reflective, and refractive objects often generate additional artifacts and noises in the reconstruction.
These non-desirable subjects and objects usually do not cause major troubles, and are automatically removed or interpolated by our algorithms, as long as they are not dominant. From time to time, It may cause reconstruction failure when they become dominant in the scene This is why you may notice that waters or rivers are not reconstructed properly.
Still photos or images are better than videos in the accuracy and the quality of 3D reconstruction.
The frames extracted from videos suffer from the artifacts and noises due to rolling shutter, motion blur and others.
It is recommended to use the photo mode of a drone to capture still photographs, not videos, even though Altizure is capable of processing the video frames!
It is recommended to use at least a 4K video camera as the camera resolution determines the reconstruction quality and accuracy.
A panorama photo is not 3D! A panorama is created by stitching overlapping images taken at the same position but rotated around. A 3D scene is created by using overlapping images taken at the different positions!
A panorama is equivalent to one single image!
It is possible to reconstruct 3D from two panorama at different positions, but these two panorama have to be overlapping, that is, see the common reconstructed area.
It is not recommended to use panorama data for 3D reconstruction purpose.
In general, there are two ways to capture photos with drones.
The circular (POI) flight. You fly a circular path around the building or the object by pointing the camera with 45 degrees, usually. If you are using a DJI drone, it is recommended to use DJI Go's POI flight to go around the buildings or the objects to capture images.
The circular POI is good and efficient for a very small project such as an invididual builidng or object.
The scanning flight. If you are using a DJI drone, it is recommended to use Altizure App (see FAQ on Altizure APP) to control your drone in this way to capture images.
The scanning is an automated systematic manner, which is good for any area to be reconstructed.
Moreover, there are two different scanning flights, one is the vertical scanning flight, and the other is the oblique scanning.
The vertical scanning flight points the camera vertically downward. If you are using a DJI drone, it is recommended to use Altizure App. The first path of the APP gives the vertical scanning of the site.
The vertical scan is sufficient for generating a 2D map or an orthomosaic.
The oblique scanning flight points the camera with 45 degrees downward. If you are using a DJI drone, it is recommended to use Altizure App. The APP provides the four oblique scans from the second to the fifth path.
The oblique scanning captures better building facades and other cancave parts of the scenes, should be used for true 3D models and more systematic capturing of large areas.
If you are using a DJI drone, it is recommended to use Altizure App (see FAQ on Altizure APP). The APP provides the complete set of five paths including one vertical scan (the first one) and four oblique scans.
Automatic capture. We have developed an APP named Altizure, both for iOS and Android, downloadable from Apple store and Google Play Store, free of charge. You only need to specify an area you want to map and the height of flight, then the APP automatically generate the fly paths and capture all necessary still photographs from different angles. Then, you upload them to Altizure for processing and enjoy the results.
Manual capture. Or, you may fly your own pattern and manually trigger the camera shutter from the remote control to capture any overlapping photographs for your need.
The automatic capture mode works well for a large area to be mapped. The latest Altizure APP enables the mission management and the continue functionalities to facilitate the large-scale capturing.
The semi-automatic and manual capture mode work well for a small area or an individual landmark building, a statue, and an object for it to be modeled in greater details.
It is recommended that you always automatically take vertical views using the Altizure APP, then use POI (Point of Interest) Mode to capture the landmarks at lower height. The vertical views above are important to make your model complete! and the other oblique views, the POI views, or other manually-shot views supplement by adding more interesting details.
The circular POI flight
In this mode, a drone flies around the object of interest and keeps its camera looking at the object.
The Circular POI flight is more suitable for capturing a single building or landmark.
It takes fewer images in this mode, while the 3D reconstruction is still at high quality.
If the building is very high, you may fly the drone by going around several times each at a different height, so to make sure that both the roof and the bottom of the building are captured in details.
The scanning flight
In this mode, a drone flies in a zig-zag flight path to scan the whole scene. It is more suitable for capturing a large areas.
One of the major measurements of resolution is Ground Sample Distance (GSD), which is in the unit of cm/pixel.
Given the same image resolution, the lower your flight height is, the higher the GSD you will get. A higher resolution of 3D models means that there are more fine details.
The following table shows the required number of images and flight height to capture one square kilometers of data with 5, 10 or 20 GSD and 70% overlapping.
|Pixels (MP)||GSD (cm/pixel)||Flight height (m)||Number of images||Total pixels (Gigapixels)|
|DJI Inspire 1 X3||1200||5||117||2037||24.4|
|Sony Alpha 7R (Lens Sony FE 35 mm f2.8 ZA，equipped with DJI Z15-A7)||3600||5||358||732||26.5|
|CANON EOS 5D MARK III (Lens CANON EF 24mm f/2.8 IS USM，equipped with DJI-5D III)||2200||5||192||1154||26|
The estimated number of images is only for your reference without any warranty. The actual number of captured images highly depends on the control of the flight trajectories, weather, the shooting interval of cameras, the shape of the captured regions, and the operation of the UAV pilot.
Image Overlap Images need to have at least 70% overlap with adjacent ones.
Camera Angle Besides taking photographs with the camera orthogonal to the ground, you will also need oblique images where the camera is tilted 30-40 degrees to the vertical direction.
Vertical and oblique photographs Usually, we are taking vertical views from a drone, that is, the camera is pointing downward to the ground. But to have more details on facades of a building, we also need to take oblique views, that is, the camera is pointing to the ground with an angle of, for instance, 45 degrees. As we need to take oblique views from all possible directions, we usually take four oblique views and one vertical view at each position, again make sure that the images of the same set of direction are overlapping between the neighboring positions.
Flight Path You will need to fly your UAV to cover the target zone in zig-zag fashion, as is shown in the figure. The actual flight path should be slightly larger than the intended data capturing area. Specifically, the extra distance should equal flight height. For example, if the flight height of your UAV is 100m, the extra flight distance will be 100m.
Photogrammetry is to make measurements from photographs.
Computer Vision is to understand images and photographs, and to automate its processing.
The end lap is the overlapping of successive photos along a flight strip.
The side lap is the overlapping of adjacent flight strips.