Support Website Contact Support Blog

Blank Area in Project

 We have been using Pix4D for about 1 year and have processed number of sites , with no issues previously. One of our recent project site is showing big blank areas on the outputs. We flew drone twice on that site, used different keypoint image scales (based on recommended settings for agricultural sites https://support.pix4d.com/hc/en-us/articles/202560159-How-to-improve-the-outputs-of-dense-vegetation-areas-#gsc.tab=0)  , and every time output products have huge blank area all over. 

Attached are few screen shots of how my project is looking like after processing. On the second screenshot, you can see a portion is hanging out at a weird angle outside the project boundary.

Any thoughts or suggestion would really help.

Hey Vikrant, 

This is probably due to uncalibrated images of the trees, since there is almost no distinction between the trees it is unable to generate key points and therefore not able to stitch them.  You mentioned you have used different image scale options, try to use 0.5 (half image) scale, which would increase your chances of getting a complete mosaic. Moreover under Initial Processing/Calibration tab, there is an option to change the Targeted Number of Key Point, as default it is automatic. However if you change that to manual and put 40,000 or something similar to that, you would also increase your chances getting a better orhomosaic. Would you be able to take screenshots of your pdf report by any chance? I may be able to help you more if I could take a look at it. Thank you.

1 Like

Hello Selim,

                  Thanks for your response. I have attached report screenshots. Does image scale in step 1) Initial Processing and step 2)  Point  Cloud and Mesh , both needs to be 1/2 image size ?

Page 1 of the Quality Report says it all…4 blocks and a RED alarm icon so you have 4 totally separate projects going on here because of lack of matches.

Solution 1: fly again with increased overlap

Solution 2: with existing pictures maybe increasing matches works, as Selim said by upping the image scale…I would go 1X or even 2X if your hardware can handle it

Hey Vikrant,

Definitely try to change the image scale 0.5 for the Initial processing, and for the step 2 you can leave it as default. Moreover the number of key points by default (Automatic) 77514 is very high however in your case you should try to manually force Pix4D to lower that number by inputting 40000 or somewhere around that.

As Adam mentioned above try to increase the overlap. You may also try to do multi level flight one flight at 5 cm resolution and the other being 2 cm resolution. For the step 3 DSM Orthomosaic and Index Details also try to force Pix4D to use 2-2.5 cm resolution for the DSM and Orthomosaic resolution (normally it would be automatic as default but change it to manual and force it to be 2 or 2.5cm). 

Basically what you are trying to achieve is get 1 solid image block, instead of 3-4 different image blocks separate from each other.

Selim, why would you say key points should be lower?  That is the opposite direction from increasing image scale for Step 1.  40,000 key points should be enough but clearly the matches are not enough so run everything higher in Step 1, not lower.

Unfortunately the best solution is simply to fly it again with an increased overlap…that is always Solution #1 for a multi-block result.

Hey Adam,

You are right higher key points and higher image scale is ideal and the lower key points may produce lower quality image, however by lowering the key points there would be less key points in overall and the relationship between the images would be stronger therefore there would higher chance of getting a stitch orhomosaic. I had a two full day training from Pix4D related to precision agriculture, and that is what I have been told then. 

1 Like

Selim,

I can see the logic in that (although it could be lowering the weak images by the same percentage and have no net benefit) but I also caution you on believing everything you hear about “how to use Pix4D” no matter where it comes from.  I haven’t tested this particular case as my work is in 3D engineering modeling but in that line of work if I had listened to everybody else on “how to” then 1mm relative accuracy on a 200’ tower would still be impossible.  Thankfully I didn’t listen to everybody saying it can’t be done (10mm to 30mm is the best accuracy possible they all said) and now that technology has caught up to me, amazing real world accuracy can be achieved for true as-built 3D BIM engineering analysis.  Innovation doesn’t keep following the current set of rules and most everything we all do here in photogrammetry is “new”.

I am just saying that while taking everybody’s advice is helpful, be sure to test it yourself because even the people at Pix4D don’t always know the full capabilities of their software…they can’t test every use case in the world :slight_smile:

 

Hey Adam, 

My interpretation is not just based on what Pix4D people told me, I have tested it with about 100 projects so far since I am extensively working with the precision agriculture data. That being said the method that I have been using in order to get stitched orthomosaic is not a %100 effective solution but in most cases it does generate a better stitched output compare to the default options. I am also well aware that Pix4D don’t always know the full capabilities of their software, there were multiple instances that I was the one who told them there were issues, which was acknowledged afterwards :) 

1 Like

Hi guys, 

Interesting discussion! I’d like to add some information about the Keypoints Image Scale at 1/2 for forest or dense vegetation areas.

In such areas extracting keypoints in individual images is generally easy. The problem comes when trying to match the keypoints between different images. As the geometry of such areas is very complex (trees, branches, leafs,…) each keypoint will represent a small detail in that complex area and finding matches will be tricky or error prone. This can result in uncalibrated cameras or noisy reconstruction.

Reducing the Keypoints Image Scale to 1/2 can be seen as reducing the resolution of the image at which the keypoints are extracted. Thereby, one pixel will represent less details of the complex geometry, which should improve the matching between images (as the keypoints will be more similar).  On the other hand, if the Keypoints Image Scale was increased, e.g. to 2, then the pixels will represent even more details, which will increase the complexity of the matching. This can be counter-intuitive at the start. 

Note that from experience we noticed that the matching for forest and dense vegetation areas is better with a GSD at about 10cm/pixel or more. In the project from this post the average GSD is at 1.38cm. Flying higher could improve the reconstruction as well. 

Selim, thanks for clarifying for me and I am going to try taking the image scale down to 1/2 on my current project that is a structure out in the middle of the ocean…quite challenging with sky and water.

I’m sure that I am wrong, but the annotation tool works very well and seems faster in processing. Especially in sky/water situations.

I like Pix4D’s idea here as this tool is used in almost all image/video editing softwares.

Down side is the tool is not well implemented and would benefit greatly from manual and automated masking options.

 

Hey Adam,

Yeah it definitely sounds confusing, but give it a shot using 1/2 image scale and lowering the key points. And as you and Pierangelo mentioned it would be very challenging to get stitched orthomosaic with the complex geometries, such as high dense vegetation, water and sky, desert because everything looks the same. 

Well I ran 1 project twice with only changing the Image Scale from 1 to 1/2.

The 1/2 Image Scale didn’t match 4 of the pictures that the 1 Image Scale did so at least in my pictures the higher resolution did better matching.  My projects are for 3D engineering modeling and not ortho surveying so it is an apples to oranges comparison.

The pictures show that the lower image scale did help reduce the overall noise but it also removed a significant section of pipe on the left side.  The overall point cloud reduction was just under 10%.

1.0 Image Scale

1/2 Image Scale

1 Like

Hey Adam,

You are totally right, it is comparing oranges to the apples, my experience is specially focused on agricultural data so I am not sure what to suggest about your 3D project, because I have no idea.

Well Selim, I am pretty happy with my results but I am always looking to improve the workflow…including trying some non-Pix4D options [gasp].  Considering the structure was in bright sunlight and in the middle of the ocean, I think the detail in the point cloud is exceptional given that the pilot only got me about 1/3 of the pictures I wanted.

But back on this thread’s topic…if there is enough overlap in the pictures then there will be no issue with Pix4D processing it even when there is nothing but sky and water in the background.  Doing nadir survey/ortho work should be even easier than my 3D modeling.

Guys,

         Thanks for your time and suggestions. Looks like we need to fly again with much higher overlap.

1 Like