Guide to generating great A.I. datasets and 3D models

1. Generating great visual datasets for A.I. neural network training

How to collect relevant visual data for A.I. training

Whatever a human child can identify with their eye, our A.I. Unleash live neural network can identify as well.

Object training via aerial images

As a rule of thumb, you need at least 500 images for each target object to identify. 5000 images are best.

The importance here is the variety.

The image shots should be taken from:

  • different distances
  • different angles
  • different lighting
  • different texture

The total resolution of the image is less important, the minimum resolution of the object should be 100 px in height and width. Ideally, you take all images with the same resolution, i.e. 1080p or 2.7k. 4k is often too much pixel information and tends to slow down the training process.

Often we combine different A.I. algorithms, for this, it is helpful to add:

  • different backgrounds
  • different sensors (e.g. thermal)
  • different dates (make sure all images are date stamped)
  • geotag captured within the image

Special case for motion flow A.I. models

Unleash live is uniquely able to conduct motion flow analysis of moving object(s) via live IP camera and drone imagery.

To capture good training data please observe the following:

Fix the camera to a pole and have it completely still, without any movement.

Hover the drone and point the camera in nadir, add image stabilisation, make sure it is completely still, without any wind or navigation stick input.

Make sure no other object moves in the video except for the moving target object (typical foreign objects moving through the video are grass or trees or water or birds, avoid these movements).

Start building your A.I. models

Once you have collected and uploaded all the images and videos via the Unleash live mobile app or via our browser interface, you can start training and building your custom A.I. models.

To learn more, read What is the A.I. Apps Ecosystem.

2. How to optimise your drone shoot for a 3D model scene

Think of it as a spray can, the more you spray !equally! across your scenery, without gaps, the better the results.

  • Fly on a bright sunny day, mid-morning or early afternoon
  • For 3D objects, we recommend DJI Point of interest shoots. Do at least 2 rounds with 20–40 images each: We recommend to shoot at 5 m/s rotation, shoot one 4k image every 3 sec. Fly one round at 25m, one round at 50m, one round at 75m.
  • For terrain models we recommend DJI Grid: We recommend shooting with at least 60% overlap, this is at 50m height, one shot every 3 seconds, with 15m/s.
  • Ensure the images are all geotagged.
  • Upload all image files to your Unleash live account directly from your tablet or laptop. We recommend 50–150 images to start.
  • Images stitch only if there is no sky in the background!! So keep at least 45-degree camera angle.
  • Select at least 20 or more images in your library to create a model.
  • Start a modelling job from within your Unleash library tab menu.
  • The model construction may take anywhere between 10 min to 3 hours. Please be patient.
  • Check the Unleash solutions page to view and explore your completed model.

What you can do with the data — Five examples:

Live object detection: https://youtu.be/ao-JfiJG3qw

Object count in images: The Unleash live A.I. engine allows for generating image-based analytics, such as counting objects. The accuracy depends on each A.I. model, our stringent quality tests usually ensure 80%+ identification accuracy.

Motion analysis: https://youtu.be/zQN1eSQZPv0

High resolution 3D Models : https://youtu.be/WZdPVsT4CkA

Similar to the library, models are stored in Solution tab folders that are named according to the device used, date and time.

Each model contains different downloadable types of files that can be used in VR, Engineering and Architecture software:

  • PLY 3D Mesh
  • OBJ, GLTF Georeferenced textured mesh
  • LAS, CSV, XYZ Georeferenced dense point clouds
  • PNG, GeoTIFF Orthophoto images
  • DEM, DTM (upon request)
  • Engineering grade Point clouds
  • Volumetrics
  • VR Experiences

Still need help? Contact Us Contact Us