Dany Bittel reconstructs blueberry with 3D Gaussian splatting

A blueberry. 90 stacks, 68 photos each. 1.48M splats. #3dgs

Gaussian splats truly is one of the coolest things that’s enabled by modern cameras and computing power.

@DanyBittel we need a library of everything at this granularity

@LinusEkenstam I'm on it. 😂

@DanyBittel That is incredible, how long til we get a full MRI style scan of someones anatomy totally made in splats

@no1089 @yacineMTB idk why but this reminded me of an old nvidia product i used to play around with when i was younger

nvidia gaugan, pretty cool

https://blogs.nvidia.com/blog/gaugan-photorealistic-landscapes-nvidia-research/

@s7tmbr Like this. Takes around 20 min to shoot.

@kvickart That would be cool.. maybe if I freeze it, then thinly slice it. That's not really splat though.

@Artemisia_Lux A focus stack. The camera moves on a rail and shoots 68 photos, each has a slightly different area of the blueberry in focus. Then I use Helicon Focus to combine the images into one fully sharp image.

@DanyBittel That’s insane detail! Would love to see this raspberry placed in a full scene.

@saianonymous Well this single raspberry is already maxing out my GeForce 3060ti. 😂 Some improvement is in the works, but I don't think it'll replace traditional rendering. The light is also baked.

90 stacks, 68 photos each.

This means at 90 different angles, OP took 68 photos at sequential depths with a macro lens and focus stacked (using something like Zerene Stacker) those 68 frames for each angle for a fully focused image of each angle. oP then used those fully focused images to construct a point cloud using photogrammetry software which then can also be used to project the images onto the point cloud. Pretty fun stuff tbh

@OIeg_Ivanov_008 @DanyBittel here's a nice video that explains where it's beeing used right now: https://www.youtube.com/watch?v=TnqK1tkREOg

@hydraulicspood @DanyBittel @kvickart ok tried with apple sharp but the problem is x-ray scans are ortographic projections and also there are not enough data to get the depth, but there was a software creating volume data from x-rays and mri's gotta try that one.

@jwt0625 Stackshot and boom arm.

@hampusolsson Thanks.. I'll post a d/l next week, so you can try yourself.

@danialghods_ If you have the COLMAP dataset, just drag and drop it into https://lichtfeld.io/ or https://github.com/ArthurBrussee/brush or https://www.jawset.com/ (probably easiest, but subscription based).

@the_Feign Sort of, but not what you have in mind. Instead of photographing you can just render each view. They are called synthetic Gaussian splats.

@danialghods_ The light is different (beauty vs flat). The interior (as well as the bottom) is not very clean though.. I'll post the splat next week.

@J_Cartterr Aak Dany!

High quality inputs is usually the answer.