Questions arising from a large task

#1 After 70 hours I had some weird/missing results with the ortho when using ORB (more on that later), I’m trying again with SIFT, with these parameters:
Created on: 18/11/2021, 15:06:45

Processing Node: node-odm-1 (auto)

Options: crop: 1, dsm: true, matcher-distance: 60, matcher-neighbors: 24, mesh-octree-depth: 12, mesh-size: 300000, min-num-features: 11000, pc-rectify: true, split: 2000, split-overlap: 200, use-3dmesh: true

When matching image pairs it appears there are 3 possible results:
1/
2021-11-20 13:19:36,821 DEBUG: Matching DJI_0304_5.JPG and DJI_0301_5.JPG. Matcher: FLANN (symmetric) T-desc: 51.936 T-robust: 0.026 T-total: 51.972 Matches: 701 Robust: 659 Success: True

2/
2021-11-20 13:21:19,594 DEBUG: Matching DJI_0020_5.JPG and DJI_0924_9.JPG. Matcher: FLANN (symmetric) T-desc: 67.559 T-robust: 0.018 T-total: 67.582 Matches: 30 Robust: 11 Success: False

or
3/
2021-11-20 09:26:11,558 DEBUG: Matching DJI_0749_6.JPG and DJI_0804_9.JPG. Matcher: FLANN (symmetric) T-desc: 54.974 Matches: FAILED

EDIT- it appears this 3rd instance happens when there is almost no overlap, which could happen when the terrain is higher, without allowance being made in the flight plan.

However, it appears some points are misplotted, not using the EXIF position.

The single plotted camera position is a long way from the images either side of it, within the dense camera area below it. The position in the header looks to be ok, so why would it be plotted so far away?

1 Like

Why matcher-distance: 60? What’s your pixel width and GSD?

60m was my guess for what I thought it should be.
Drone is an M2P, GSD averages around 22mm, but varies a bit over rugged terrain.

In the first attempt at this large dataset, I tried using ORB and the camera.json file from a large subset of the same area, and it took 70 hours, but missed large areas around the edges. All of this area has previously successfully been combined into smaller orthomosaics.
The first attempt settings:
Options: cameras: {“hasselblad l1d-20c 2048 1365 brown 0.7777 rgb”:{“projection_type”:“brown”,“width”:2048,“height”:1365,“focal_x”:0.7777778798476017,“focal_y”:0.7777778798476017,“c_x”:-1.4353584987681075e-9,“c_y”:2.7706085708877087e-8,“k1”:4.1401891459771804e-7,“k2”:2.502759404555409e-7,“p1”:0.0000016387400806644321,“p2”:5.352177439969789e-7,“k3”:1.504782081263804e-7}}, crop: 0, debug: true, dem-decimation: 3, dem-resolution: 5, dtm: true, feature-type: orb, matcher-distance: 60, matcher-neighbors: 20, mesh-octree-depth: 12, mesh-size: 300000, min-num-features: 15000, orthophoto-resolution: 5, pc-rectify: true, use-3dmesh: true, verbose: true
Average GSD: 8.22 cm
Area: 8,448,784.18 m²
Reconstructed Points: 91,682,871


This was resized by 0.33X

The report pdf is here: https://drive.google.com/file/d/1AUveHikWgUbERRQXHz_PA_2HBMv4LXNl/view?usp=sharing

As can be seen there are massive GPS errors, which there wasn’t with the smaller subsets, and the DTM has a vast extra area to the south and east which was nowhere near the area imaged, and some elevation points way below reality too, with the point cloud having points all over the shop, which explains why the DTM is so far out. But why is the question?!

I’m currently over 70 hours into a 2nd attempt (with a long way to go) using SIFT and split/merge, settings as per original post, and resized by 0.5X this time, in case too many features were lost in the downsizing.

1 Like

Just thought it was a bit high.

Doesn’t resizing change the effective GSD

Are the missing part in the ortophoto over trees?

Yes resizing changes it, doesn’t change the matching distance though. Your .ods calculator gives exactly 60m for height matching distance for my 70/65 overlaps.

No, much of the missing area is open ground with a few trees and plenty of well resolved rocks - plenty of features, and they did combine well when I did smaller areas.

When you did the flight, did the drone stop for every photo? I’m thinking it might be some blurry photos.

No time to stop with nearly 1500hectares to cover! I’ve looked through all >23000 photos and they are all sharp and well exposed, although the image quality does drop off around the edges in lower light conditions when the aperture was wider than f/5, particularly the top left corner with the Hasselblad camera (same in all 3 M2Ps I have used). I monitored exposure and shutter speed while flying, making sure exposures were short enough for flying speed to avoid motion blur. As mentioned earlier, subsets of this area combined well.

Can you try matching-distance for width and increase matching-neighbours to what my ods say?

1 Like

I can try increasing from 60 to 83m matcher-distance, but it would be a reduction from 24 to 20 on the neighbours on the re-run of the task currently in progress. 75 hours into it now, with a long way to go before trying anything else.

1 Like

You can only have 20 good matches anyway, the corner ones is cut away in my calc. But to get all 20 with matching-distance you need 83m.

1 Like

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.