all for open source, but these are two different products.
from their website:
> To fully utilize Meshroom, a NVIDIA CUDA-enabled GPU is recommended. The binaries are built with CUDA-10 and are compatible with compute capability 3.0 to 7.5. Without a supported NVIDIA GPU, only "Draft Meshing" can be used for 3D reconstruction.
Somewhat off topic from the article, but I'm curious about GPU vs CPU computing here:
So it looks like meshroom disables a feature (or features) when not having a Nvidia CUDA-enabled GPU, rather than e.g. still have the feature, but have it be slower.
Any idea, if it were computed (or emulated) on CPU, how much slower it would be? (like 10x slower, 100x slower, 1000x slower, even more?!)
And also, ignoring speed but instead looking at implementation effort, if it's feasible to emulate on CPU at all, or is GPU really so different for this that it's too much work to implement the same thing for a general purpose processor? At least matrix multiplication libraries (the main thing GPUs can do well afaik) exist for CPU too...
From my experience with deep learning I can eyeball it: one or two orders of magnitude. The fine details may be different on this particular problem.
In DL in general, if you have sort of equivalent CPU/GPUs (think a 5600 vs GeForce 3070) you'll find yourself in the range of 10 to 30x faster training with GPU. If you have substantially different platforms, you can start looking at 100x speedups or more.
For example, my mobile 4600H CPU trains about 100x slower than a desktop GeForce 2060. And close to 300x faster than the old Intel 7200u I use as a "slow system" to test things. Which also has a GeForce 940MX, that trains about 30x slower than the 2060 (if you manage to make your model small enough to fit in 2GB of VRAM).
Is it not embarrassing that viewing PDF is not just baked Into the OS? As a Mac and iOS user it boggles the mind when I see colleagues on Android closing ad popups so they can view a PDF. Why are such things accepted?
There's a native reader in the Drive app. It's a nice integration actually, allowing for quick management of the file in Drive, keeping it from vanishing forever in my phone's "downloads" folder
I'd been worried about that when I first tried it too. But it opens giant (130MB+) PDFs on my SD card just fine when in airplane mode and when I've had no cell signal. So despite the name, the Drive PDF Viewer app seems to run locally.
That's good to know, but still leaves me with the uneasy feeling of if/when this will change. For my taste, the cognitive load of all these security issues on mobile is a bit too high.
Yeah, I am sure they also upload the metadata of every PDF file you save / open in Acrobat (I am sure Apple Books also does this) and I just find the whole Apple iDevice ecosystem irritatingly intrusive as hell for this reason - they deliberately do not have an application firewall for Wifi that would plug all such data leak.
That said, there are some use cases where Acrobat may need such permission for their "pro" version:
- Contacts: For sharing and / or collaborating on a PDF file.
- Photos / Media / Files: For editing or authoring a PDF file.
- Camera: For converting a photo of a document into a PDF file. For inserting a photo (e.g. signature) into a PDF file.
- Wi-Fi connection: For syncing / storing a PDF file into their "cloud".
But even if they do actually need these permissions, it should still work as a "Reader" / viewer without these permissions and / or demand permission only when a feature needs those permission, instead of a blanket opt-in.
FYI it is in fact possible to manually adjust app permissions after you install them and Android will block usage if it's ever attempted. I usually axe permissions for basically everything and only enable what's actually related to core functionality.
Does that mean that Whatsapp can't ring my phone when it's locked? Yes. Do I care? No.
Recently found out that my Moto G4 phone is running an Android version of 4 years ago. It didn't run any security updates since, and without warning. And I can't upgrade. Phone hardware is perfectly fine otherwise.
This isn't really directly Google's fault. hardware vendors refuse to support bsps for older hardware, much of which is closed source binaries that Google cannot maintain. You might notice the support window for pixel hardware that doesn't use Qualcomm socs is significantly longer.
Both Chrome and Firefox have built in sandboxed readers. Just drop a PDF in them. Chrome on iOS also reads PDFs. My last android died but I'm pretty sure it's the same.
Have you thought about the product requirements? It makes a lot of sense because many people's best camera is their phone and the app can give them real-time guidance to make sure they took the picture correctly.
I think the disconnect here is that you seem to think the step-by-step is a tutorial on how to use AliceVision, it isn't. It's a high level explanation of how the photogrammetry pipeline works. If it helps, you can think of it like a pop-sci article covering scientific research. If you're looking for a tutorial on how to use the tooling built around this pipeline, you can start with [0], found by clicking Meshroom in the menu and then the tutorial link.
I tried installing Meshroom and it felt pretty rough, but it turns out I had simply failed to restart my Linux box after installing the CUDA toolbox. The software seems stable, though I'm still messing around trying to better understand it. My initial attempts with the default settings and a handful of photos of my mouse could use some work, and it's pretty slow on my box (an aging Xeon and about the oldest graphics card that's compatible with CUDA for this software). But it's interesting stuff, and better polished than a lot of research software I've worked with.
I guess you're not in the "If you can" category then, seems basic reading comprehension is a fading art. I for one appreciate that the tutorial attempts to help you understand how the process works and why. Perhaps it's just my enormous brain at work but I found that when I opened the program up the "drag your images here" space and "start" button were pretty self-explanatory.
(please read this with dripping sarcasm and imagine me rolling my eyes very hard)
On a more serious note I really think meshroom has done a great job encapsulating a complicated process in a way that is simple to execute on the outset and yet affords the end user a great deal of control over what's going on. I really like the way they visualize the process with the node graph and that they cache the results at every step so you only ever have to recompute changes downstream of the node you've edited. I love tools that let me solve problems the toolmakers didn't specifically envision.
I'll take the attempted insult. But at no point do you approach anything that would fall under the definition of "sarcasm", unless you're trying to agree with me. You're so far away from even mere irony that there will probably be rain or your wedding day (don't remember the song).
Call me a cynic (it's in the dictionary), but I used to be insulted by smarter people. I shall have to think about that.
Anyway... I've actually done photogrammetry (with the Apple API– wish I could share my excellent rubber ducky, but Wikimedia Commons doesn't quite like my file format). I was mostly just pointing out that I had different expectations from a "step-by-step guide", and that OP's wish of OSS adoption might be more effectively supported by making them want to do so, rather than shaming them with slogans from the mid-90s.
Sometimes, a more powerful product is more complicated. But the same people who will eagerly post "correlation is not causation" on every science story as if that weren't the first thing any scientist learnt will happily pretend that something being difficult to use is evidence of and necessary for a more "professional" product.
You're right that I'd definitely be doing more good if I wasn't being an dick about it, sorry. Though probably I shouldn't have engaged anyway. However, I am still drinking so despite my better judgement let's get into it on the facts of the matter at hand. (dang please forgive me for degrading the quality of HN)
> mostly just pointing out that I had different expectations from a "step-by-step guide"
Let's not pretend you were doing your best to communicate this in good faith with your original comment, but I shouldn't have been insulting, again, sorry.
However you absolutely touched a nerve because as someone who's made it a point the last few years to try to switch to using only OSS Meshroom really does a comparatively bangup job of usability.
> Yes, but the step-by-step guide has eight steps.
Sure the guide could be written better, but it does a good job of explaining the process and a great deal of related minutia.
> And the first step cites five papers.
Come on now, if this isn't meant to paint it as some sort of inapproachable academic process (which is not the case at all) why mention it.
> But at no point (at whatever depth) does it tell me what to do.
Step Three absolutely clearly(ish) tells you what to do, and that step alone is enough to get a decent result. Yeah the formatting could be better, and yes there could be a quick start guide/TLDR that says "take a bunch of photos from a lot of angles without moving the object, drag them in and hit start" but come on, if you read through it it definitely does tell you what to do, hence the reading comprehension comment.
Regarding sarcasm, certainly at least the bit about "my enormous brain" fits the definition.
I've been using 3D Scanner App [0] for almost two years now. And it has been free and very feature rich this whole time. So many very helpful export options, as well as a nice frontend. I've also made use of getting the raw data here and there. Its hard for me to tell what this adds over that. And like another comment points out, this looks like a thin wrapper over the tools already provided by Apple.
The App Store page says "Simply take photos of an object from all angles and upload them to our service, and we will create and send you a 3D model ready for 3D printing, Augmented Reality and our web app r3Dent." so not it appears to use some kind of server-side processing. So it's not even on-device.
Looks like the same tech we used at https://usdz.app . It will be very interesting to see what new APIs Apple brings out this WWDC. Hopefully the photogrammetry API makes its way from macOS to iOS.
In the mean time, if someone wanted to make their own app, they could use: https://usdz.app/api
I want one of these 3D photo apps but geared for object replication rather than art projects and videogames. Like let me scan an object then add measurements to it that let me more perfectly recreate it, such as identifying curves for me to input a measured radius, asking whether this section is supposed to be a straight edge or a flat plain or a perfect cylinder.
My experience with professional solutions a few years ago was that it worked well for large areas but the tolerances weren’t anywhere near good enough.
A “trick” seems to be that it looks great with the photos overlayed as textures, but the actual 3D data without that isn’t nearly as impressive.
Exactly. I feel like AI could do a great job with this by starting with a rough scan and then asking the user intelligent questions about what certain shapes and measurements are.
Take photos in a circle around object sounds suspiciously like object vr. Usefull but old. But it also suggests it may be using an iPhones lidar face sensors to scan the object which would be quite cool. There are no obvious examples though? and I don't have an iPhone with face recognition so who knows. Rather vague on prices to export models?
Seems like it's a hot topic and there are a lot of clones lately in the App Store. And probably Nvidia after its latest demo is disrupting the market again.
The only one that I think offers a cool demo/use case is Holograms.
https://alicevision.org/#meshroom