Photo alignment Issue: Strange Behavior with first Photo Alignment step

[AerialGlobeSSE]

I am looking for help trying to understand why my imagery is not aligning in a predictable manner and what workflow procedure could be recommended to resolve this issue. 

The dataset is of a short and straight two lane road with no trees (Approx 1 km long oriented East/West), collected from a car mounted extension with GoPro Camera (Linear mode), looking down at pavement at slightly oblique angle. Two videos were collected - Eastbound, then Westbound at 30 km/hr, then Imagery was extracted from video every .10 seconds for proper overlap.  I have successfully collected and processed imagery this way many times with the same settings (No GPS). 

The align step (default settings) only succeeds in aligning approx. 60% of the images (1200 of 1900 images). All successfully aligned images are from the Western end of the road, then the alignment drops and the tie points end abruptly at a certain point as you approach the eastern end of the road (even though all Eastern end images are there in the dataset). 
I did try different accuracy settings “High” and “Highest” – increasing the tie/key point limit and it did improve it slightly but still had the abrupt drop of alignment
I began trying different things to resolve the issue and this is when things got strange.  I tried splitting the dataset up into two. One folder of the EVEN numbered photos (from both east and west runs - photo 2,4,6 etc) and a second folder of the ODD numbered photos (from both east and west runs - photo 1,3,5 etc) My thought was I will take every other photo and process them as two chunks and see if I had better luck merging the chunks.  I noticed that when processing the smaller chunks I was still only getting approx. 60 percent success on the alignment step and they were also geographically situated on opposite ends of the road.   (See Screenshot)

The dataset of ODD images aligned approx 60 percent of the images on the western end of the road, and dropped approx. the last 1/3 of the eastern portion (same results of full dataset above). but the dataset of EVEN images aligned approx 60 percent of the images on the eastern end of the road and dropped the last 1/3 of the western portion!

I was beginning to think that maybe there was something wrong with the collection of the imagery on the eastern portion of the road – not enough overlap - and that’s why it was not aligning, but when I ran the EVEN dataset, the Eastern portion aligned perfectly so that hypothesis does not seem to hold.   

I am not sure why splitting them into even / odd datasets causes either the Eastern OR Western end to align but I would appreciate to hear anyone’s feedback or suggestion how to improve the overall alignment since I have determined seems possible.

I recently did try to align the chunks and merge them, but I couldn’t get that to work on my first attempt.  I wonder if I am just missing something here.  (Align Chunks then Merge Chunks)

Please reply with any questions.
Thanks !

I included a image of the Eastbound image thumbnails so you can see how much overlap I am getting on the Eastern end of the dataset .  Looks like plenty of overlap (between 3-5 images with the same match point contained) and that is just in the Eastbound dataset.  The Westbound dataset has the same overlap,  so probably i have between 6-10 match points at any given point.

[Paulo]

hey Aerial Globe,

30 km/h with image extracted every .1 sec will give you overlap overkill...about an image every .83 m.... I would try aligning only one image in 3... or so and see...

Do you have the specs for your Gopro camera (width, height, focal, pixel size)? as well as height over road and obliqueness of camera shots?

[AerialGlobeSSE]

Thanks Paulo so much for the suggestion ~ I had not thought of overlap overkill!
I did try every third photo (1,4,8,12 etc.) but unfortunately the results were considerably worse off with the same alignment settings (Accuracy: High, Key point limit 40k Tie point limit 6k).  It only aligned 7 of 628 images with this arrangement and photo selection.

The settings for the GoPro Image Collection are:
4:3 aspect ratio at 2.7k resolution which is 2704 x 2028 pixels.
The height of the camera off the ground is approx. 3.5 meters high with an approximate oblique angle of somewhere between 35-45 degrees from NADIR.
I have spent the past few months experimenting and testing my camera mount set up and have been able to achieve results creating quality orthomosaics with this same configuration. 

The main logical issue I don’t understand is why does one half of the imagery (odd number dataset) render good alignment on roughly half the road, but not the other half of road ~ but when I flip it and process the other half of the imagery (even number dataset), the half that would not align in the odd dataset now will align, and then the half that was aligned with the odd dataset, will not align with the even dataset? 

If this were an issue with some sort of collection or image overlap issue I was thinking I should expect to see possibly more of a “random” distribution of the misaligned photos throughout the entire dataset, some here, some there being misaligned ~  but the fact that it is not picking up any alignment for any image the entire other half of the road ( in either dataset) that is what I just cant make sense of, when that portion does not align in one dataset, but it will in the other.
 
(I am able to verify that the collection speed was consistent throughout both runs.)
Any other questions or suggestions are appreciated ~ Thanks!

[Paulo]

Maybe share a data set for internal testing? if possible....

[AerialGlobeSSE]

Good Idea ~ Thanks Paulo!

I would be happy to share the dataset - Can anyone recommend how i can share it?
(Approx 2000 photos ~ 480MB)
Thanks

[Paulo]

Just use a web sharing service, Google Drive, Dropbox,....

[AerialGlobeSSE]

Thank you for the recommendation.
The imagery can be downloaded here:
https://drive.google.com/drive/folders/1Cp5wSpLrHQ9u2tGRwdi-UIIM_rJoY6W5?usp=sharing

Reach out with any questions.