The wall can be made of 4 vertices, and it is made of 10,000 splats. Therefore, the best compression method will find objects that are flat and change them into a formula, things that are round become spline formulas, which is like isosurface research from the previous few years. It's a project for a bigger lab like Nvidia or AMD. A kind of hybrid research between splats and iso3d. :D
Nice. These tools should allow for saving the ply file with the level of SH of your choice. Perhaps, in some cases, using only up to level 2 SH could be good enough, case in which almost 30% of the data (the SH3) would be discarded, resulting in a much lower file size.
There is actually an interesting automatic optimization feature if you use the Polycam service and copy the Embed code to put the GS model into your own webpage. I noticed that when using the Embed code, Polycam drops the model's SH degrees by 1 or 2, depending on how fast the embed loads on your pages. I think it's a well thought out and creative way to utilize model optimization.
What is the size of the original poligon model? It is good to see a comparison between the polygon format and 3d Gaussian splatting from all aspects. Kind regards
Thank you for all your videos. I'm very passionate about gaussian splatting, I could even say addicted hehe. Thank you for the tips about sharing on Polycam. For me, the best way to watch GP is in virtual reality with Gracia. I was interested about the compression because in VR, sometimes the framerate is not very good. I would like to have a way to combine many GP to do a complexe scene. I made many tests to see what are the best options for capturing : DSLR pictures, raw pictures with my phone or UHD videos with the phone in manual or auto. Each of them has pros and cons. DSLR give very crisp GP, have a very large dynamic range, large depth of field, control on the image and I can make sure not to move during the pictures. But, it takes longer to capture the images so I often have less angles than a video and it takes longer to prepare the images. The videos give me many angles, very fast and easy, but I need to have a very short shutter speed to reduce the motion blur. It would be nice to have a video comparing the different capture methods and what is best between a few high quality pictures or many lower quality pictures ( video ). I also noticed that sometimes, it's better to capture many pictures from a restricted angle than moving all around the object. Like a normal picture, sometimes the composition is good only from one point of view.
Great. Experiments like that with photos and videos with different shooting tools are really good experiments that help to find which method produces the best looking Gaussian Splatting model. I encourage you to keep experimenting. One interesting way is to also try producing synthetic Gaussian models. Through them, you can simulate several scenarios that are not always possible to produce with 3D scanning of the real world.
@OlliHuttunen78 Yes I really like your tool. It's a great idea. I tried it before by rendering a video in Twinmotion to see my model on LumaAi. I also experimented with funny tricks with Gaussian Splating. For example, you can do a 3D scan of someone's head, but the person is following the camera with his eyes without moving the head. When you watch the GS, the person is always looking at you ! You can also have half the the person scan with a expression and the other half with another expression. I also make some objects disappear on certain angles by removing the object when I was capturing this angle.
@OlliHuttunen78 Yes I really like your tool. It's a great idea. I tried it before by rendering a video in Twinmotion to see my model on LumaAi. I also experimented with funny tricks with Gaussia Splating. For example, you can do a 3D scan of someone's head, but the person is following the camera with his eyes without moving the head. When you watch the GS, the person is always looking at you ! The person can aslo change his facial expression for half of the capture angles. I also make some objects disappear on certain angles by removing the objects when I was capturing this angle.
Thanks Olli! Still not a fan of subscriptions. If you happen to stumble across a service that let's you publicly share (and OWN!) your models with a perpetual license on the web, please let us know! Really interesting video though 👌
Hi Ollie, great video as always, Two things came to my mind when I saw it. What about the .splat format, isn’t that also a bit smaller than the PLY files? I think some of the viewers that are based on 3JS do support .splat files… when comparing the MCMC format to the previous version of postshot, I find that these files are a lot smaller as well . 280mb instead of 750mb - I know MCMC usually produces better quality, but I found that that is not always true, but I don’t know why. I assume it’s a little bit like video compression one is compressed with variable bitrate and produces smaller files while the other one is compressed with fixed bitrate. Maybe you can look into this in a future video.
.splat files are basically same as the compressed PLY files. Especially if the .splat file is produced through the Super Splat web application. File size can be very small but Spherical harmonics data is still lost in that format. MCMC is based more on random sampling versus the older ADC method which uses more linear ways to produce gaussians based on gradiants. Both method have they pros and cons.
Hi Olli, have you found a good ply importer/viewer for Blender? I got 2 different importers working, but the results were very bad and Cycles didnt work.
There is only two at the moment. One that Alex Carlier released in Github some time ago. It works only with cycles. Second one is bit better that Kiri engine guys developed. It uses Evee engine. I’m still waiting that we could get splats running natively in blender like it is already possible run them in game engines. Perhaps some day.
Hello! I watch a lot of your videos and I don't understand why they have so few views. I'm an engineer and I think that with gaussian splat you can take measurements of buildings and other things. I do this, but no one can help me! Now I have one main question, on the luma ai website you can download a point cloud with colored dots, but if you make a gaussian splat in postshot, then there is no such option! I don't understand how you can convert gaussian splat to point cloud while preserving the colors of the dots, to work with this!
Seems like nobody has bothered making an efficient file format for these whatsoever. From my reading, ply is poor from the get-go in how it stores numbers. Quickest way to maybe get some results would maybe be to simply zlib compress the files, it would certainly require the least effort. Lol.
Thank You Ollie 👏
The wall can be made of 4 vertices, and it is made of 10,000 splats. Therefore, the best compression method will find objects that are flat and change them into a formula, things that are round become spline formulas, which is like isosurface research from the previous few years. It's a project for a bigger lab like Nvidia or AMD. A kind of hybrid research between splats and iso3d. :D
This video was exactly what I was looking for. Thank you so much Olli!
very informative clip! kiitos!
Nice. These tools should allow for saving the ply file with the level of SH of your choice. Perhaps, in some cases, using only up to level 2 SH could be good enough, case in which almost 30% of the data (the SH3) would be discarded, resulting in a much lower file size.
There is actually an interesting automatic optimization feature if you use the Polycam service and copy the Embed code to put the GS model into your own webpage. I noticed that when using the Embed code, Polycam drops the model's SH degrees by 1 or 2, depending on how fast the embed loads on your pages. I think it's a well thought out and creative way to utilize model optimization.
Thanks a lot.for your research on GS
your videos are incredible, thank you so much
I love your channel! Have you examined Gaussian Splat models made from HDR images?
What is the size of the original poligon model?
It is good to see a comparison between the polygon format and 3d Gaussian splatting from all aspects.
Kind regards
once again a very useful video !!
Thank you for all your videos. I'm very passionate about gaussian splatting, I could even say addicted hehe. Thank you for the tips about sharing on Polycam. For me, the best way to watch GP is in virtual reality with Gracia. I was interested about the compression because in VR, sometimes the framerate is not very good.
I would like to have a way to combine many GP to do a complexe scene.
I made many tests to see what are the best options for capturing : DSLR pictures, raw pictures with my phone or UHD videos with the phone in manual or auto. Each of them has pros and cons. DSLR give very crisp GP, have a very large dynamic range, large depth of field, control on the image and I can make sure not to move during the pictures. But, it takes longer to capture the images so I often have less angles than a video and it takes longer to prepare the images. The videos give me many angles, very fast and easy, but I need to have a very short shutter speed to reduce the motion blur. It would be nice to have a video comparing the different capture methods and what is best between a few high quality pictures or many lower quality pictures ( video ). I also noticed that sometimes, it's better to capture many pictures from a restricted angle than moving all around the object. Like a normal picture, sometimes the composition is good only from one point of view.
Great. Experiments like that with photos and videos with different shooting tools are really good experiments that help to find which method produces the best looking Gaussian Splatting model. I encourage you to keep experimenting. One interesting way is to also try producing synthetic Gaussian models. Through them, you can simulate several scenarios that are not always possible to produce with 3D scanning of the real world.
@OlliHuttunen78 Yes I really like your tool. It's a great idea. I tried it before by rendering a video in Twinmotion to see my model on LumaAi. I also experimented with funny tricks with Gaussian Splating. For example, you can do a 3D scan of someone's head, but the person is following the camera with his eyes without moving the head. When you watch the GS, the person is always looking at you ! You can also have half the the person scan with a expression and the other half with another expression. I also make some objects disappear on certain angles by removing the object when I was capturing this angle.
@OlliHuttunen78 Yes I really like your tool. It's a great idea. I tried it before by rendering a video in Twinmotion to see my model on LumaAi. I also experimented with funny tricks with Gaussia Splating. For example, you can do a 3D scan of someone's head, but the person is following the camera with his eyes without moving the head. When you watch the GS, the person is always looking at you ! The person can aslo change his facial expression for half of the capture angles. I also make some objects disappear on certain angles by removing the objects when I was capturing this angle.
thank you
Thanks Olli!
Still not a fan of subscriptions. If you happen to stumble across a service that let's you publicly share (and OWN!) your models with a perpetual license on the web, please let us know!
Really interesting video though 👌
Hi Ollie, great video as always,
Two things came to my mind when I saw it. What about the .splat format, isn’t that also a bit smaller than the PLY files? I think some of the viewers that are based on 3JS do support .splat files… when comparing the MCMC format to the previous version of postshot, I find that these files are a lot smaller as well . 280mb instead of 750mb - I know MCMC usually produces better quality, but I found that that is not always true, but I don’t know why. I assume it’s a little bit like video compression one is compressed with variable bitrate and produces smaller files while the other one is compressed with fixed bitrate. Maybe you can look into this in a future video.
.splat files are basically same as the compressed PLY files. Especially if the .splat file is produced through the Super Splat web application. File size can be very small but Spherical harmonics data is still lost in that format. MCMC is based more on random sampling versus the older ADC method which uses more linear ways to produce gaussians based on gradiants. Both method have they pros and cons.
This question might be off topic. In your opinion, do you get higher quality models from using Photos, or from Videos in postshot?
Playcanvas are working on retaining the SH in future.
Hi Olli, have you found a good ply importer/viewer for Blender? I got 2 different importers working, but the results were very bad and Cycles didnt work.
There is only two at the moment. One that Alex Carlier released in Github some time ago. It works only with cycles. Second one is bit better that Kiri engine guys developed. It uses Evee engine. I’m still waiting that we could get splats running natively in blender like it is already possible run them in game engines. Perhaps some day.
Hello! I watch a lot of your videos and I don't understand why they have so few views. I'm an engineer and I think that with gaussian splat you can take measurements of buildings and other things. I do this, but no one can help me! Now I have one main question, on the luma ai website you can download a point cloud with colored dots, but if you make a gaussian splat in postshot, then there is no such option! I don't understand how you can convert gaussian splat to point cloud while preserving the colors of the dots, to work with this!
Seems like nobody has bothered making an efficient file format for these whatsoever.
From my reading, ply is poor from the get-go in how it stores numbers.
Quickest way to maybe get some results would maybe be to simply zlib compress the files, it would certainly require the least effort. Lol.
bs
Reducing the file sizes of Gaussians? If you’re referring to a specific part then please put a time stamp.