Hi all,
many of you who are new dont yet have the experience of what makes photos GOOD for photogrammetry ( for converting to 3D)
we made this new video to help out
we also have the Getting Started Guide that ends with a written section about it http://remake.autodesk.com/assets/Autodesk_ReMake_Guide_01.pdf
finally we still have the old Webinar 03 in which we tried to explain all that in more detail. https://youtu.be/SBuH223QZoc
I found one of our really old photos was used in the video. And we weren't informed.
Hi,
I have a little figurine that I was trying to 3D capture. I think I did everything right but my object is not turning out that great so I am starting to think I am missing something. Here's what I did:
- lighting setup: mounted two fluorescent light balasts (with two lights each) vertically, facing away from the object.
- kept the object stationary resting on top of a pattern surface
- took 100 photos around the object at three different planes
-camera setup: fixed light balance, appeture 22, using a fixed lense (Canon EF 24 -70mm, fixed at 70mm), remote control, mounted on a tripod.
Please help? I can send the 3D capture and photos if that helps with the troubleshooting.
Thanks
Mark
Hello meng1313,
I'm happy to help you out. Nothing in your description is a red flag, so the best way forward would be to share the photos, or at least a few photos so I can analyze what is happening.
You can either post a link to a Dropbox-like folder containing your photos to this forum so everyone can learn along with you, or if your photos are confidential then send me a link privately: b r i a n -dot- m a t h e w s /at\ Autodesk -dot- com. Note there is only one letter "t" in my last name. Alternatively if you don't have a service like Dropbox you can just email around 3 of your images directly. But the more images I have the more likely I'll be able to explain why you didn't get a good result.
Note that your aperture is a little small. You want the smallest aperture you can get such that everything is in sharp focus, but for most cases anything smaller than f/19 is going to create diffraction. Check out http://www.kenrockwell.com/tech/focus.htm
-Brian Mathews
Autodesk
Hi Brian,
Thanks for the prompt response!
I will provide a link to the files in Google Drive which is here for your reference: https://drive.google.com/drive/folders/0BzRYtNu3gsX0dkhZcEllUDRLQkE?usp=sharing
I will try it again using optimal sharpness.
-Mark
Mark,
Thank-you for the sample images. There are several things I can point out:
1) Check out image IMG_3532.CR2. Note that until the entire process is done the computer doesn't have any 3D and thus doesn't know what is foreground and what is background. The computer is going to process *every* pixel in every image. Everything needs to be in focus because the first step in the process is to determine (a) where the camera was in space -- called a camera 'pose' estimation, and (b) what the lens being used is and the lens distortion. To do this we find corresponding features between images. Even the features in the background contribute to helping the software understand where the photo was taken from. Notice the chair in the background in one photo gives your brain a clue when you see the same chair in the background in another photo -- your brain knows the positions of those two shots were near each other. The software does something similar.
In summary: everything needs to be in sharp focus.
2) Lack of unique features. In that same photo, notice what your object is sitting on: a regular grid. When you look at a pixel where two green squares intersect on the left side of the table it looks identical to the two green squares that interest on the right side of the table. When the computer looks at a different image (say IMG_2533), it sees a two-green square and it needs to find a correspondence to the prior image. But since all green squares in the first image look the same, it can't find a good match-- the probability functions don't collapse to one solution. The computer has a hard time figuring out what lens you have and where you were standing.
You'd have far better results putting your object on concrete where there are a lot of organic and random patterns at different sizes and scales. Some people use newpaper, etc.
Also look at Joker-3-55.jpg -- in that photo the background is devoid of features. The only features are of the object itself. If you watch our training video you'll note that there are cases where you want a featureless background (when using a turntable) but if your background doesn't have good features then your foreground had better have them. Which brings us to the next issue:
3) Man-made Objects have few features. Your joker is made of featureless monochrome plastics. Within his arm it is solid purple. It is very hard to match one spot on his sleeve in one picture to the exact same spot on his sleeve in another picture because there aren't any relocatable marks. The computer can sometimes see features the human eye can't and experienced users can still get this to work by sparying the object with a splotchy powder that creates unique surface features. If you had a better in-focus and random-pattern feature-rich background it might allow this object to work even though it has low surface texture except for the next issue:
4) Your object is shiny. That is a killer. The computer is matching features in the view-dependent shine reflections and those reflections are the surface itself. Look at joker-3-59 and how the left lower jacket has a shine on it, but that same location in picture joker-3-58 does not have shine in that location. The computer can't find a match because that part of his sportcoat isn't the same between the two pictures. For experts we can solve that with circular polarization and special lighting. Your best bet is to go outside on a cloudy day such that you have highly diffuse light with no shadows. But that may not be enough.
So in summary you've selected a difficult object with a difficult background with a difficult focus issue. Two of my 4 suggestions are easy for a novice, but two of them might only be fixable by an experienced user. You can become experienced but it takes practice and understanding of how the computer does its magic and how filters, powders, polarizers, lighting, focus, framing, and features can be added.
Good luck!
Brian Mathews
Hi,
I watched some the best practices demonstration videos on youtube and have recently tried to take photos of a rock using the 'turntable' method. I have made a couple of attempts which have been unsuccessful and wondering if anyone would be able to help me identify what I am doing wrong/missing.
I have the files saved in a Google Drive folder.
Thanks in advance
-Mark
Sure: where is the link to the Google Drive folder?
Without seeing the photos, I'm going to guess that you would have had better results if you didn't use the turntable method. While the turntable seems easy in concept, in practice it is hard to make the background featureless - the computer can see what you can't with your eye and the computer doesn't understand the difference between background pixels and foreground, so if some of the features are rotating and others are not then you get bad results.
I suspect you'll get much better results if you keep everything (including the lights) stationary with the camera being the only exception. Also place the object on a background that has a lot of random patterns such as concrete. Diffuse light helps too. The best results are often outdoors on a cloudy day (no shadows, full diffuse light) with your object on concrete (lots of random multi-resolution features).
-Brian Mathews
Autodesk
Hi Brian,
The link to the Google Drive folder is:
https://drive.google.com/open?id=0BzRYtNu3gsX0WGtpNVJWNVBXVXc
You are correct - I had better success with the stationary method (rock-4) vs. the turntable method (rock-2) and both are saved in the folder in the link above.
Based on what you said in your previous message, it seems that my background has too much texture which is being detected by the software. I'm currently using a large piece of white paper to remove any creases, however I assume that the software is able to pickup the texture of the paper as well which may be causing some of the issues? If this is the case, would you be able to recommend an appropriate background and lighting setup?
Currently I am using two sets of two fluorescent light tubes (40W) each, mounted vertically on tripods.
-Mark
OK, some comments on your two tries:
ROCK-2 -
ROCK-4
-Brian Mathews
Autodesk
Dear All,
I had a look to your video which learns us the best practices to take photos.
I took only 31 photos because I try to learn for coming projects and make a POC (Proof Of Concept)
When creating 3D model from my photos, at the end, I have a popup saying that there was a failure "Local Reconstruction Failed!"
Here is the share drive where I have copied my photos, and as well as the log from Autodesk Remake Product Version 17.25.0.16
https://1drv.ms/f/s!Ams36kCLXHYEkcc-8xHGmpX9cHh6BQ
Could you please help me on this and tell me where the problem comes from, since I don't undesrtand the log file.txt.
Many thanks for your advices,
Regards,
Francois Lepron
If you read my post at the top of the thread you will see that you have not really followed my tips for getting a good capture.
Specifically:
1) Your photos are not in focus. Every pixel needs to be in focus, not just the pixels covering your object. The software finds features and matches them across photos (not just the pixels for your object, but also background pixels). Consider your pictures and how the software will have a hard time making a 3D model from this:
My guess is that you are using a Cell Phone camera. Don't. Get a real camera with a real quality lens. Use a tripod to keep motion blur to a minimum. Use a smaller aperture but not so small that you get diffraction effects.
2) Your object is only covering a tiny portion of the pixels. It looks like 1/3rd horizontally and 1/3rd vertically which means your object is about 10% of the pixels. You are throwing away most of the resolution of your camera.
3) The background behind your object is a sea of red pixels. There are no visual and random shaped features for the software to lock onto. I dare you to take a pixel of that red surface in one picture and find the exact match of that spot in another picture. If the human visual cortex can't match a feature, a computer can have a hard time too. Instead place your object on a randomly textured (but not shiny) surface like concrete or newspaper.
4) Don't use local compute. The system requirements are very, very high. It is only for experts who want to spend a lot of money on very powerful computers. Use the cloud while you are a beginner.
5) You need more pictures. You want a picture every 5 to 10 degrees of angle. That means 36 in a loop at one layer, then take some additional in an arc going over the top of the object.
Good luck and have fun practicing. Taking pictures seems so easy because you've done it your whole life, but doing photogrammetry is not as simple as it seems. You have to understand how it works and really follow the guidelines exactly as shared.
-Brian Mathews
Autodesk
Many thanks, Brian.
I will follow you advices and let you know what are the outcomes.
Regards,
Francois.
Regarding the scan of your circuit board: I should have added that even if you do it correctly, that circuit board will still have issues. If you follow the tips you'll get a better 3D model, but a circuit board will never come out great for a number of reasons. For example, it is mainly flat to begin with, the chips on the surface are black and featureless, there are shiny parts like the USB connectors that have mirror reflections, there are many occlusions (the connector pins for example), and the scale of the features you are modelling are very small and fine (1 to 2mm) and as such you start to get depth-of-field focus issues if you get close enough to see the detail. Photogrammetry works best with organic objects that have a lot of random surface detail, and less well with monochromatic / shiny / featureless / smooth man made objects.
If you really want a circuit board scanned with high precision and accuracy you should use a different type of technology such as a laser line scanner. Here is one example of such a device from one company (there are others): https://www.youtube.com/watch?v=0Hyp0SDY9sA&feature=youtu.be Our ReCap product can work with laser scan point cloud data from such devices. They aren't cheap.
-Brian Mathews
Autodesk
Hello
I am new to photogrammetry and I am experimenting a little bit now. I would like to know if there is a way to take photos of an object (e.g. a bowl) from the bottom and use them with those from the other sides to make a model. I tried twice but it didn't work because there are either too many "common" points or no connections. I do not want to make 2 models and combine them later.
Thank you.
It is possible to capture both sides of an object in one project without having to stitch two geometries, but it is tricky for beginners.
If you read my other posts you'll see that the system looks at every pixel and in the first stage it is trying to locate the position of the cameras and determine what lens is being used (called Pose Estimation). The software will use both pixels that include your object of interest, plus all the pixels of the background surrounding your object. Normally the more unique "visual features" you have the better -- including the background pixels. This is why I suggest people put their object on something like concrete or newspaper since those background pixels will help the software determine the pose and lens distortion parameters more accurately with more points to work with.
However, if you want to do a double-sided capture then you need to do the opposite of my normal recommendation: you want to make sure that there are no "visual features" in the background. The idea is to suspend your object on a featureless background (and not have shadows either!) and take a bunch of photos, then flip the object 180 and take photos of the other side. Since there are no features on the background it won't get modeled by the software. Consider the "light box" shots that people use to make products stand out on a white background for eBay product shots -- they use special diffuse lighting and a curved featureless (and fully diffuse / non-shiny) white material behind the object.
In practice this is much harder than it sounds for beginners. They often don't understand what diffuse light really is, or diffuse surfaces, and don't really provide a true featureless background. Remember that a computer and camera can see details that your human eye can not see. The computer is perfectly happy to match against those features. Then things like shadows can become an issue because the shadows move relative to the object when you flip the object over. Those shadows create visual features the software will try and match against.
You can also do some Photoshop work to try and remove the background and force it to 100% white for all background pixels. But again, lighting and/or shadows on the object itself may be an issue.
Finally, if you remove all the features from the background then the question is what features does the software use for Pose Estimation? It only has the features of your object to go on. If you object doesn't fill the frame then there may not be enough features to match against for a quality pose result. Or if you object is relatively low on features (like monochromatic plastic, etc.) or if your object is shiny (mirror reflections that change based on camera angle and thus not consistent from shot to shot), then your chance of success is low.
One solution is to use a ring-flash for illumination (normally I say not to use a flash since it creates moving shadows, but a ring flash is different) with special circular polarizers to try and remove glare and shine. That costs money and requires other photography knowledge beyond this post.
So, yes, we do what you are suggesting all the time and get great results. But we have been doing this a long time. Most beginners will read this and not have success. The upside is that if you have the right equipment (light box, diffuse lights, ring flash, polarizer, fixed focus lens, DSLR with high resolution, etc.) trying to follow these guidelines won't hurt anybody!
-Brian Mathews
Autodesk
Hello.
It's a mystery for me. I made a lot of try. The last one is not fail, but the files mesh is totally idiot !
I took 111 pictures of my object.
I choiced normal mode in my cloud.
What's wrong in my way ?
Sincerely.
Can't find what you're looking for? Ask the community or share your knowledge.