Hotspot Effects in Micro Copter Data

A shots from last year vegetation point cloud mapping in Brandenburg. Strong Hotspot-Effect visible with easily visible shadow projection of the quadrocopter shape on the tree crown texture. Nice example of this direct back-reflection effect on vegetation (crown surface in this example), although you can also clearly see the same effect on the grassland vegetation in the neighborhood of the tree crown.
smh-20170527-d4c5abaf23a147c434e47d2a0105ac83-0759

 

 

Absolut Accuracy of the Phantom 4 Pro Ortho Mosaic

Some results from DGPS point measurements with a Stonex S9III GNSS DGPS receiver indicate an offset of 1.5-3 m to the Phantom 4 Pro ortho mosaic images. The relative accuracy is however much better. The Phantom 4 Pro derived ortho mosaic (without reference points) has a standard deviation of 0.31 to 0.36 m.

When comparing the overall positional accuracy of the reference points its surprising that the UTM coordinates that come directly from the Phantom 4 Pro are not far off: I measured 12 DGPS points and found that the offset is between 150 cm and 290cm (compare with Fig. 6).

RMSE between the Stonex S9III DGPS (SAPOS) measured points and the DJI Phantom 4 pro POI mode ortho image mosaic measured points is 1,98 m (stdev 0,31 m) and the RMSE between the Stonex Points and P4P NADIR mode ortho image mosaic measured points is 2,95 m (stdev 0,36 m). The difference is clearly also depending on overall atmosphere conditions and Stonex DGPS accuracy. 

 

dgps-point-comparison-12pts

 

Fig. Comparison of measured DGPS Stonex GNSS points with in-Copter derived GPS geometry in POI ortho mosaic and with Nadir ortho mosaic processed in Agisoft in September 2017.

Speeding up Agisoft Photoscan

For some of the copter data processing folks Agisoft Photoscan turns out to be the most important tool to calculate point clouds, orthoimages and nadir data mosaics. Problem: very long processing times with dense point cloud calculations with high or ultra-high settings (full resolution image matching with SfM (Structure from Motion) algorithms).

Some nice net finds show how multicore processing has its limits and why you should invest into GPU performance … and in high end 3D graphic hardware.

Combining more than 20 CPU cores doesnt seem to speed up the process and combining more than 4 GPU systems also doesnt seem to help. There is only a minimal speed increase when you add more CPUs and/or more GPUs when a 24 core system is already installed.

http://www.nvidia.com/download/driverResults.aspx/115283/en-us
http://www.agisoft.com/downloads/system-requirements/
https://www.pugetsystems.com/labs/articles/Agisoft-PhotoScan-GPU-Acceleration-710/

It boils down to a dedicated system with 2-4 Graphic cards with 3D acceleration (GTX-1080ti cards from Nvidea), with approx 64-128 GB RAM and a dual i7 system setup.

mtk – Sören