Cost Effective Results

Aiming to produce high quality output is a persistent goal here at AccuPixel. We use high end DSLR cameras for both terrestrial and underwater work and the drone is very much intended to produce images suitable for high accuracy photogrammetry.

But at the same time we are aware that image quality is rising whilst the cost of equipment falls. Using “cheap” kit will always require caution and validation but what if we can work within the limits to produce output that is fit for purpose?

A road scene subject to forensic investigation captured with GoPro and processed in Metashape Professional
A road scene captured for collision investigation using a GoPro – would it hold scale?

Sample Data

We were contacted by Chris Goddard of Collision Science and offered a couple of sample datasets created by low cost kit – A GoPro action camera and DJI Mavic Mini drone.

On the face of it, neither camera would be near top of a “recommended” list. Both have inherent design features that are not really suited to deliver accurate photogrammetry.

But would they deliver a result? And would the advantages of recursive optimisation and scaling in Metashape Pro refine the output and could we hold a tolerable level of accuracy?


We are big fans of camera poles. Raising any camera increases coverage and can help with overlap, as well as the obvious advantage of covering anything that is beyond arm length.

The scene had been walked with the GoPro on a pole, shooting 463 images in four minutes. After initial alignment and initial camera optimisation the distance cones laid into the scene were used to apply scaling – we would not recommend using short distances to constrain a large model – but they were fixed/known values and were better than nothing.

After final optimisation we added a check bar – to confirm distance rather than constrain the model – and variance over 20m was just 20mm.

During image capture cars – moving objects – had transited the scene. The ortho photo needed a little patching work to remove the transitory elements. This would be expected irrespective of the capture tool and with over 400 images to choose from finding suitable replacements was easy.

The resulting ortho photo (one of the outputs available from Metashape Pro) was reporting a scale of 2.49mm per pixel, a very good level of detail. The 1.2Gb tiff output isn’t directly viewable on the web, but our friends at Dronelab have sorted that. Here’s the ortho photo – be sure to zoom in:

At certain points the camera pole appears as a repeating feature. This is almost certainly down to technique and is a straightforward correction. Again, the ortho photo patching tool can be used to remove them.

During the investigation Chris used the point cloud output from Metashape to analyse the incident in PC-Crash:

Metashape dense point cloud was used to provide context and background to the collision analysis. Screenshot courtesy Chris Goddard/Collision Science

Overall, we would have preferred to see constraints over the entire scene, and better still if accurate ground control points were placed and measured with GPS to give geo location. Care needs to be taken when working quickly with small CMOS sensor cameras and the quality of resulting images must be preserved. In this well lit scene the GoPro had little difficulty achieving low noise blur free images.

In the conditions the results were impressive and outputs more than fit for purpose.

Mavic Mini

A short flight had captured 142 images of a road junction.

All drone images will contain location data – latitude, longitude and altitude – embedded in the jpeg files. We were not expecting accurate GPS data from the Mavic Mini and were not surprised to find accuracy anything other than “thereabouts”. There was no way the GPS values gathered could be used for reconstruction.

After initial alignment optimisation started to refine alignment. Although GPS was of little value we can still apply constraints, adding scale bars before final optimisation, using the markers placed into the scene to fix known dimensions.

The road junction model derived from Mavic Mini images

Again, using a small CMOS sensor we must be mindful of light levels, ensuring quality of images are consistently high for photogrammetry. And like the GoPro scene adding control points would have delivered accuracy over greater distances and geo location.

Scale bar errors remained within an acceptable limit – just 90mm over 50m.

The ortho photo scale was not quite the same level as the GoPro derived images, delivering 12.6mm per pixel and this could be further improved if higher resolution was required. Hosting of the ortho photo again is by Dronelab:

For all its limitations the Mavic Mini has one advantage – its takeoff weight puts it in the lowest risk drone category. This means it can be flown in areas where heavier drones could not be deployed without significant risk mitigation in place.


Neither model could be geo located or referenced but local scaling was within reasonable and usable tolerances. Were we surprised? Yes, but then again we do frequently remind ourselves and clients of a simple mantra when capturing and processing 3D data;

Will the outputs be fit for purpose and meet client requirements?

This guiding principle helps guide us as to what equipment to use and the processing settings applied. There is little value in buying hardware costing thousands when the client needs (and budget) can be delivered by sub £1k kit, or by setting Metashape processing parameters at highest and wasting processing time. We cover all these points in far greater detail during our Agisoft endorsed online Metashape training courses.

One positive benefit is keeping time at-scene to a minimum. A laser scanner would have taken far longer than the four minutes the GoPro walk around took.

Would we use this kit for every instance?

No, not at all. But if the client requirements fit then cost effective hardware – used within its recognised limits – will deliver.