No announcement yet.

More than 90% of images rejected....

  • Filter
  • Time
  • Show
Clear All
new posts

  • More than 90% of images rejected....

    I snapped about 350 photos of a potential home building site with survey markers in the ground and 3DFlow rejected more than 300 of them... I think it was because there was too much sunlight in the photos but I was hoping someone could look at the photos and give any other advice on how to improve the import success. I shot at high aperture and 200 iso. Thanks for any suggestions.

  • #2
    ThursdayJuly23rd - Thanks for sharing the images! - Downloading now and will give some feedback.


    • #3
      Thank you!


      • #4
        Hi again - So I ran your photos through with default settings and got similar results to you, but by increasing the matching depth in the advanced settings I was able to get 335/359 images aligned.


        Click image for larger version

Views:	119
Size:	39.4 KB
ID:	6624

        - Then the default dense cloud came out really nicely:

        Click image for larger version

Name:	dense.PNG
Views:	125
Size:	1.41 MB
ID:	6623

        Having two or three different height passes would be beneficial as currently, you are shooting all your images from the same level, also looping in circles around the main subjects should greatly improve the chances of successful camera orientation without the need to use the higher processing settings.


        • #5
          Thanks very much for experimenting/helping on this model. I'm happy to know that the settings can be tweaked to get more matching depth.

          I went back to the site today and re-shot everything. This time I:
          Took 704 images ( before 359 )
          Used 18mm zoom level ( crop sensor ) ( before used 10mm on portrait mode )
          F16 ( before F22 )
          Shot at 3456x2304 coarse ( before 5184x3456 fine )
          Increased overlap.
          Added 3 or 4 panorama points.
          Circled more of he open areas.
          Shot on a total overcast day ( before was sunny )
          I still only shot from one level today; took 4.5 hours to take 704 pics.

          I'll try processing the model overnight and report back.

          Thank you!
          Last edited by ThursdayJuly23rd; 2020-09-15, 05:25 AM.


          • #6
            Nice! Look forward to hearing about the results.


            • #7
              Thanks again for your assistance on this model.. my 2nd attempt at photographing the site got 466 out of 704 images aligned ( using your suggested match depth settings ) . Here is a quick video I made up showing how the model turned out ( it did not turn out good at all ) :

              Here are the images I used:

              Thank you again for any tips on how to get better results.


              • #8
                Hi ThursdayJuly23rd - Your second set of photographs they appear to be scaled down to 3456*2304px from 3456*5184px - Do you have full resolution versions of these images? They might have been scaled down in Lightroom accidentally, or in-camera.

                Prior to input into Zephyr, make sure not to crop or resize the images and don't apply any distortion correction (distortion correction is calculated for within 3DF Zephyr) - Keeping the JPG files at 100% quality will yield the best results and minimize artifacts/camera misorientation.

                It looks like you might be shooting photographs from the same position in space and rotating the camera - you want each of your photographs to have a unique position in the scene to get better coverage and co-ordinate information for camera orientation. Also, try to keep a uniform distance between photographs, rather than having clusters of tight-knit shots - if you went looping and supplemented with a broader pass to ensure each cluster is associated with the next that might work well!

                I think also the 10mm is better suited than the 18mm as you get a broader field of view and have a better chance of feature overlap between shots.

                Rather than trying to capture the whole thing in one go, I'd approach capturing this scene in an incremental fashion, first perfecting one tree, then perfecting two trees connected with the floor etc... when you're confident with your shooting style you can then snap the whole thing up in one go.


                • #9
                  Thank you for the quick reply on this.. I'm going to try running the model using the original images from the camera tonight.

                  ========= Capture in One Go vs Incremental ===========

                  I had pretty good results on my other models; so I figured that I was ready to photograph the whole thing; it seems that getting one tree or area is easy; but capturing the photos that give 3DFlow the data needed to connect all of the areas together is the difficult part. That is why I created the other post asking about the best path to take for an area like this.

                  As you can see in my images I orbited each tree and area of interest trying to keep good overlap. I'd really appreciate some additional ideas on how I could have better covered this area.

                  ============= Same Position in Space =========

                  I did shoot some panoramic sequences but the vast majority of the images were not panoramic.
                  The 3DFlow photo tutorial recommends a few "panoramic sequences" for urban/square/hall type areas... I thought having these sequences would help tie the different areas together. So if I remove those images it would help the model? Does it hurt to have these in the model?

                  ============== Lens Related =============

                  I shot 10mm on the first visit to the site but then noticed the 3DFlow photo tutorial recommends no less than 18mm for crop sensor cameras. My other models where I shot at 10mm did look better. Maybe the tutorial should clarify when it is ok to shoot wider angles?

                  ========== Lightroom & Image Size Related==============

                  Here are the original images

                  I used Lightroom to pull up shadows and pull down highlights a little; but no resizing of image was done in Lightroom. I also checked the Lightroom exported images and did not apply any Lens Corrections. Edit: I did notice that some of the exif data was different on the lightroom exported images.

                  I'm shooting Canon 700D ( 18 megapixel - 5184 × 3456 max rez ) and using the Medium Coarse image quality/size setting... I guess any setting other than the maximum size of the camera would technically be resized by the camera before writing the the memory card. I only choose the medium setting because the 3DFlow photo tutorial recommends a bigger pixel size. Can you please clarify what size and format you recommnd I shoot for this camera? See screenshot:
                  Click image for larger version  Name:	2020-09-15_19-22-19.jpg Views:	0 Size:	152.3 KB ID:	6637
                  Last edited by ThursdayJuly23rd; 2020-09-16, 10:35 PM.


                  • #10
                    DanielMuirhead gave some great feedback on shooting the ground - It's definitely a challenge doing such a large area because you have to try and make sure every surface is captured by a couple of different camera positions.

                    Panoramic style capture works better with interiors than open spaces because you get a lot more overlapping features within a confined space, however, these photos don't contribute massively to the 3D reconstruction and are usually there for the purpose of improving the texture projection from particular viewpoints. - Panos wouldn't be overly harmful to have in your scene, but not as useful to 3D reconstruction as having photographs taken from different coordinates.

                    Not all lenses are made equally, and the lens suggestions we make are very generalized - Figuring out the perfect combination of settings and setup for a given situation requires iterative testing and is dependant on things like the weather conditions, accessibility, time constraints etc..

                    For reference, I typically use a 14mm or 24mm lens on a full-frame sensor for interiors and outdoor spaces, 50mm for texture scanning, and 50mm or 70mm for my studio turntable work. 3DF Zephyr is versatile in that you can feed it images from almost any camera setup and get results, figuring out what works best for you is a bit of a process but once you have nailed it down you will find it much easier to survey a scene/subject and know exactly what will work best.

                    Ah yes looking back I can see now that the images weren't resized, but you used a compressed and smaller JPG - You want to avoid JPG compression entirely because of lossy compression artifacts. If you're going to take your images into lightroom to edit them prior to plugging into Zephyr I strongly recommend shooting in RAW - this way you typically have a wider dynamic range for pulling detail out of the shadows and highlights and can control things like the white-balance using a color chart. Once edited you can then save out of lightroom as JPG 100% at full resolution.

                    Given ideal conditions and a high resolution (36mpx +), it may make sense to reduce your resolution as there are diminishing returns there - But because your camera sensor is only 18mpx it's already on the low-resolution end of the scale, and lowering it further before running through Zephyr could be detrimental to camera orientation and 3d reconstruction, so I'd advise using the full resolution of your sensor in this case (especially seeing as the full resolution images were more successful than the reduced JPG's.

                    With your setup, I'd shoot the 700D with 10mm on a tripod on a cloudy but still day. I'd shoot in landscape to get more lateral overlap, maybe as low as F/11 based on - And F/11 there's still plenty of depth to your shots and it lets more light in, allowing us to speed up the shutter speed and reducing the ISO - Ideally, you want to be in full manual so you have control over the shutter speed and subsequently keep the exposure consistent. ISO typically doesn't want to be much higher than 400 and shutter speed as fast as possible given the light conditions so as to capture the scene before it changes (weather/sun coming through cloud etc...) I'd capture full-resolution RAW images and use a colour chart to white balance all the images and bump the shadows up and highlights down a bit before exporting as JPG for processing in Zephyr.

                    Last edited by cam3d; 2020-09-17, 07:26 AM.


                    • #11
                      Thanks very much for these tips. I'm in the process of building the mesh using the original images; 633 of 704 images were aligned. I'll share how it looks in the morning.


                      • #12
                        Here is how the model turned out:

                        Question: Would this be an area better shot with a 360 camera? I have the Ricoh theta V that I could try to use if you think it would turn out better.

                        With regards to the best point cloud / dense point cloud generation.. would you recommend these settings?
                        Click image for larger version

Name:	zy97Orn.png
Views:	110
Size:	80.6 KB
ID:	6647


                        • #13
                          ThursdayJuly23rd - I'd avoid using 360 cameras unless you have no other alternative as they have a lot of distortion and are typically lower resolution than standard stills from a DSLR.

                          If you're still testing out your camera positions then I'd stick with default otherwise you'll be waiting a long time to see the results. I'd increase the settings once I had a really solid camera orientation in the initial processing stage.