Real-Time Processing Showcase (Live Demo)

Talk · Felix Zhou · 3DISE 2025 · 25 April 2025 · 10:30 - 10:40 · 21 min

Chapters

What the talk is about

Felix Choo, founder of Tiki3D, presents the company's Infinite photogrammetry software and processing hardware, built for city-scale mass production of 3D meshes and Gaussian splats. Felix shows aerial triangulation of hundreds of thousands of images on a single desktop PC, auto-referencing that removes repeated ground control points for recurring drone surveys, and a live demo in which a drone dataset is streamed and processed in real time during the talk.

Key ideas

  • According to Felix, Infinite aligned 142,000 images in under seven hours and 647,000 images in about 62 hours on a single desktop with 64 GB of RAM, by splitting and fusing sub-blocks automatically.
  • With auto-referencing, ground control points are needed only for the first dataset; later drone flights of the same construction or mining site align automatically to it with pixel-level accuracy.
  • Running several processing engines on the same machine lets Tiki3D squeeze more out of the hardware; Felix cites a client facility of more than 215 nodes running over 800 engines.
  • Gaussian splatting shares the photogrammetry workflow and captures palm trees, glass and balconies that are usually missing or distorted in a mesh, with output to PLY, splat and 3D Tiles.
  • In the live demo, more than 500 drone images of a European farm were streamed to a Tiki3D processing box and turned into a full-resolution mesh during the talk.
“from day one, it's designed for mass production with quality and productivity in mind.”
— Felix Zhou ·
“with our, um, new auto-referencing, uh, algorithm, you only do GCP once for all, for the very first ground-based, um, dataset.”
— Felix Zhou ·
“These are thin wire structures that usually is either missing or, uh, distorted in a, um, aesthetic, uh, 3D model like a, like a mesh.”
— Felix Zhou ·

Questions answered

How many images can Tiki3D Infinite align on one computer?

According to Felix Choo, Infinite aligned 142,000 images in about 6 hours 50 minutes and 647,000 images in 62 hours on a single desktop with 64 GB of RAM.

Do repeated drone surveys of the same site need new ground control points?

Not with Tiki3D's auto-referencing, according to Felix Choo: ground control is done once for the first dataset, and later datasets align automatically to it with pixel-level accuracy.

Does Tiki3D support Gaussian splatting?

Yes. Felix Choo showed that Tiki3D produces Gaussian splats with the same alignment workflow as its meshes, scaled across multiple machines and exported as PLY, splat or 3D Tiles.

Full transcriptabout 11 min readShowHide

Transcribed automatically (ElevenLabs Scribe) and lightly edited. It may contain small errors. Click a time to play the video from there.

Felix Zhou: Please, let's give it up for Felix. Felix, join me over here. Hello, good morning. Thank you. Enjoy.

Felix Zhou: Thank you. So while, uh, my team, Jackie, is setting up something that usually you see a, a magician is doing teasing, our processing box, and, uh, I ask, uh, our guests in the front line to be a, a witness of how this is processing un- un- online. So first, um, I'm Felix from Tiki3D. New name to, um, the, the audience, and, uh, I believe, uh, we, we, we are here, um, in, in Tokyo last year, but I, I believe a lot of people, um, uh, uh... I meet so many new people here.

Felix Zhou: Um, I was raising and keeping my hands until the last last one. In, in the morning, I meet, I meet definitely way more than ten people at, at this venue. So, um, who are we? Well, I am no new guest to the European market, and, um, I have many friends here. Um, first, I'm born in Shanghai from China and then immigrant to New Zealand and took my, uh, degree in New Zealand and got my master in Australia, where, uh, you should easily pick up from my accent.

Felix Zhou: G'day, mate. And then I moved back to, uh, Hong Kong and, um, um, there are some story with reality capturing. So I'm been in the, uh, reality capturing game since twenty ten. Um, I was talking to some early adopter or, um, software vendors, um, in, uh, including, say, uh, Saab Dynamics, while the technology was called Rapid Mapping. And then, um, I'm the guy bring Acute3D to Asia.

Felix Zhou: And, um, in twenty fifteen, it got, uh, acquired by Bentley, and I become a, uh, regional, uh, industrial sales director for the product. And founded, uh, Tiki3D since twenty eighteen. Um, while what I see what's lacking in the market is the focusing on a large scale, municipality scale project that's really utilizing the 3D mesh, really capturing, um, technology in, in the massive scale. And, um, um, sharpen up the skill for six, seven years now, I think we are ready, um, to show what we got, um, to the EMEA European market. So this is what we got.

Felix Zhou: Um, so our product line, uh, comes across processing software and hardware, capturing and some industry application. I'll, I'll go through several of them. Um, because of the short timeframe, um... Ooh, the shadow actually, um, spoil all the, all the information. So, um, we believe we're a, uh, the, the regional leader in, uh, reality 3D modeling software and, um, from where we're coming from.

Felix Zhou: And, um, from day one, it's designed for mass production with quality and productivity in mind. So first step, AT. Um, so we give a, a, a, um, our branding is Infinite. So Infinite AT is like we process aligning hundreds of thousands of images, uh, in one go. You don't have to manu-manually split them.

Felix Zhou: Um, and the, uh, our latest versions, it can be processed within four hours for the alignments with a, a single normal desktop PC. Uh, reconstruction, uh, the algorithm is the, the architect is multi-engine, uh, grid computing, uh, architect. So it's, uh, similar to a lot of the, uh, the other, um, softwares. But the, the differentiator is we allow multiple instance of engine running on the same hardware, which squeezing out all the processing capabilities even further. So, um, in, um, from some customer reviews, we can pump up the productivity as much as three hundred, over three hundred percent compared to some, uh, um, existing solution they are using.

Felix Zhou: Um, I believe a lot of, uh, our audience, uh, uh, guest has been introducing, they are capturing not only airborne surveying, uh, method, but also close up, uh, technology. Um, so, um, multi-dimensional, multi-resolution, or even airborne, terrestrial, handhold, all kind of different, um, equipment. So we, um, have an, a special, uh, algorithm in our, our triangulation mode for, um, for mixing and automatic fusion of the, um, the aerial, terrestrial, you name it, different kind of, um, um, hardware. And also, uh, we have a separate module that process stereo satellite image into a, um, 3D mesh. Um, automation.

Felix Zhou: Um, when you're talking about mass production, customer don't want to sit in front of com- their computers twenty-four/seven. So a lot of automations need to be, um, carry out. Um, GCP, it's always the pain point. And, um, we, uh, overcome the, uh, the, the hurdles, um, by automatic picking up the, uh, the targets and so on. Um, auto-referencing AD, that's something I will introduce in, uh, in a few seconds.

Felix Zhou: Um, and, uh, a lot more. So AT or alignment is the first step of all the processing. And why AT? Uh, I use a, uh, video clip to show how much pressure our clients actually put to us by a madness of way of capturing from all different angles, all different resolution. They want the ground level, um, in centimeter ac- uh, resolution.

Felix Zhou: So this is really something that actually puts a lot of pressure for, um, traditional photogrammetry software. Um, and that's, um, we come in and, um, doing a lot of tweaks in the, uh, algorithm and optimizations for that. So robustness, efficiency, and scalable. That's what we introduce our super ca-capacity mode. It's one of the four, um, aero-triangulation mode in our software.

Felix Zhou: Um, this year, we're introducing the, uh, the second generation in our, uh, uh, twenty-twenty-five version. So it's, um, set up a, a record-breaking auto sub-blocking and fusion algorithm. So it's really pushed the boundary of, say, millions of images going through the, uh, the AT alignment process on itself. Um, and the, um, the accuracy, of course, is consistent, uh, consistent across the, uh, sub-blocks. And, um, just by a normal desktop, uh, with sixty-four gig of RAM, it can process, uh, over, uh, millions of images.

Felix Zhou: So couple of showcase, uh, and I'm showing the, uh, the hardware spec that we, uh, are using in the test case. Um, hundred and forty-two thousand images, throw them in, load them in all together, and let the software run by itself. It's automatically segmentations into small chunks and then automatically fusing them, uh, with the, um, um, consistent, uh, consistent of accuracy. So the total processing time for the aero-triangulation is six hours and fifty minutes. Um, I don't know, um, your experience of processing a similar dataset, how long it would take.

Felix Zhou: Um, from our feedback, this is mind-blowing. And in, uh, in the six hour fifty-two minutes, feature extractions is an hour forty-five minutes, and bundle adjustment is three hour and a half. Um, that's what we got a feedback since never has been achieved before. Um, what about bigger ones? So, um, six hundred and forty-seven thousand images, again, put through a-- the same machine, sixty-four gig of RAM.

Felix Zhou: Um, of course, it takes longer because of the, the total number of, um, uh, images processed and, um, it's a, uh, it's, it's, it's pushing, uh, the, uh, the, uh, bundle adjustment another scale. But again, with the single machine, sixty-two hours, um, we've got a job done, and with the, um, RMS zero point seven. All right, so we covered the, uh, the scaling. Um, another thing that has been demand a lot, especially useful in the cases of, um, repetitive scanning on the same site, for example, construction site, a mining site that, um, especially these days with docking station in place, that people got drone flying over the site, um, every week, every couple of days, um, want the, uh, the data to be updated. So...

Felix Zhou: A-and also there are AI or, uh, change detection algorithm, which are relying on the, the, um, the, the pinpoint accuracy of alignment for the, um, for several datasets. So, and, um, some of the, um, project we experienced, and, and they have some small structures changed, um, progress in the, um, uh, in, in, in the timeframe. And, uh, it's not big enough or not substantial enough to be able to create some GCP for this. Um, of course, we talk-- we, we have the tools of, say, RTK drones and, um, everything, but, um, um, drone-based RTKs still give you around, say, um, plus or minus two centimeter accuracy. While, um, if you require higher, uh, accuracy, uh, based on that, mm, it's difficult.

Felix Zhou: So with our, um, new auto-referencing, uh, algorithm, you only do GCP once for all, for the very first ground-based, um, dataset. Everything update, uh, uh, dataset capturing, you never need to worry about GCP. It's auto-referencing to the previous dataset and a-auto-alignment, not only the dataset by itself, but also the referencing, um, datasets. So, um, while we're talking about GCP, even for that dataset, we're just having thousands of thousands of images, literally, there is money cost involved. So people can, um, um, do hundreds, uh, GCPs, hundred fifty, two hundred, because that's time and, uh, labor-intensive work.

Felix Zhou: But with the, uh, the referencing, um, feature, we actually have thousands of features from existing datasets, which is makes the, the, the, uh, the matching even more reliable. And, um, we can provide Pixel-level accuracy for the alignment between the datasets. Sometimes the pixels ha-has higher resolution compared to what the drone RTK can achieve these days. So this is a typical use case, um, for a some construction site with the old dataset which, which is already, uh, still inbuilt, including road. And the customer of course want a, um, updated dataset with a new constructions in place.

Felix Zhou: With that kind of scenario, you never need to worry about controlling or, um, GCPs anymore. Um, you just fly over and it's automatically, uh, aligned. Okay. So enough for, uh, aero-triangulation. Reconstruction.

Felix Zhou: So we believe, um, we are in the first tier, um, quality. Of course, people want, um, market want a good quality, uh, 3D reconstruction. Uh, it's difficult to tell whose is really, um, good compared to the other one. And, um, uh, I'm dare enough to come to Europe to say say this, but, um, um, I would say in, uh, from the, uh, from the user feedback experience, we are sitting in a, in a comfortable position that, um, we don't have a, um, um, major or I mean, s- uh, uh, substantial, um, difference comparing to, um, geometry as well as the texturing qualities, uh, to our outputs. Uh, multi-node computing, that's, um, what we, um, that's, um, a lot of software actually using these days of grid computing and scale out the projects into, uh, well, ten, 20 or, um, hundreds.

Felix Zhou: From our market sector, the largest production facility we built for our client, private cloud, um, is, uh, more than two hundred fifteen processing nodes in the same processing facility. Um, and as I mentioned, we allow multi-instance processing engine running on the same hardware, which pumps up the number of engine running in the network from two hundred fifty physical machines to more than eight hundred running engines in the network, which gives us more than three times, uh, more productivity from the same hardware and the same dataset. Of course, the, uh, the new, um, fancy outputs, 3D Gaussian splatting, we also support that. So here are three, um, efficiency test cases that, um, we have. Um, uh, for a datasets of, um, say a few thousand images, we pump up to, um, if you're looking at the processing time, two hours, two hours twenty minutes, uh, about four hours.

Felix Zhou: So the, um, the efficiency of processing, um, based on the, um, just a few nodes of the, uh, um, day-to-day machine. Um, and if you considering, say big gigapixel-wise, we are be able to achieve per node, per day, three hundred comfortably, comfortably above three hundred, almost four hundred gigapixels per day. And, um, square kilometer-wise, uh, remember we are act- talking about three centimeter or two point five centimeter resolution, and, um, we are be able to pump up to, um, three... That one's sli- uh, slightly less, uh, almost three kilometers per node per day. So imagine you put that into scale.

Felix Zhou: So I use a small video clip to, um, prove that, um, our, our production quality. So this is a typical eastern town. Quiet, classic, with a golden Buddha sitting across the river and looking over the town. That's a typical mesh you can get these days, but we didn't stop there. Gaussian splatting.

Felix Zhou: So this is a video clip that shows a typical scene that you usually get not so satisfactory result from a mesh. So you see all these palm trees, leaves, and the see-through glasses and balconies. These are thin wire structures that usually is either missing or, uh, distorted in a, um, aesthetic, uh, 3D model like a, like a mesh. But they are represented perfectly in 3D Gaussian. And we share the same workflow of the photogrammetry part, which is alignment and, um, um, aero-triangulation and the dense point and correspond point cloud and dense point cloud.

Felix Zhou: And we can scale that up into, um, grid computing to, um, push the data to be processed in multiple, uh, machines. And we also have the, um, uh, same old-fashioned algorithm to split into tilings and auto merging. Uh, and in the end, we ou-ou-output the, the major, uh, formats, PLY, splat, 3D tiles, and then we also, um, have our first generation of LED technology in the, in the pipeline. So enough for the software. Uh, if you remember in our hierarchy, we also take, um, provide the, the, the processing hardware side.

Felix Zhou: The reason being is we're talking about mass production, and we're not talking about, you know, uh, um, sometimes a lo- uh, the, the customers have limited, um, IT infrastructure in place. And, um, as a lot of guests, uh, already mentioned, when you're talking about hundreds of thousands of image, that's completely the game changer if you only have experience to process a few thousand, you know, um, twenty thousand images. So, um, here's the big boy toys for all the RSC community. So we coming from building up the, uh, the GPU dense rack servers to mobile processing carts and, um, hand carry size of the all-in-one processing box and the tri-screen. And, um, I don't-- uh, I'm not gonna go into detail here.

Felix Zhou: Um, so we have a booth there, so if you're interested, feel free to come over and have a quick look. So the last slide, I actually gonna reveal what we are actually doing here, is real-time streaming processing. So this is thanks to our partner Srigon. Uh, we actually got a dataset, more than five hundred images of a European farm. As soon as, um, I start my presentation, the, the, the computer is running a simulation of streaming the da- uh, the, the, uh, the images from the drone to the box and got processed in real time.

Felix Zhou: And now by the time I finish my process, the, uh, the data's, uh, the data are already completed. So for, uh, any of you that don't, uh, can't see the screen, I have a recorded footage that's showing how fast we are be able to process. Um, and, um, remember, we are streaming the, uh, the original resolution of the full image and got processed and output of the mesh is still haven't been, um, shrink down. So, um, yeah, we have a booth there, so if any of you interested, please come over. Um, thank you very much for your time.

Felix Zhou: Perfect. Ladies and gentlemen, let's give it up for Felix.

Speaker

FZ

Felix ZhouLinkedIn

Tiki3D

With 25+ years of global experience in tech and 3D innovation at industry leaders like Microsoft, Samsung, and Bentley, I bridge IT expertise with Reality 3D solutions to tackle real-world challenges and efficient data production needs. My latest venture, Tiki3D, delivers cuttinng-edge software and…

Hear the next ones live – 3DISE 2027, 20–21 April, Prague

Buy tickets