Vincent

Building a Visual Cortex for AI & Robotics.

[Vincent @joqivi]

I will submit Liv to app store by Friday July 31, 2026 or I will pay $50 at https://t.co/jUDiiU449K (via @levelsio)

[Vincent @joqivi]

Tracking in a dynamic environment sucks.
Sky moves
Ocean moves
Humans everywhere
Cars move
#colmap become unreliable.
Masking moving elements kinda solves it.
Lesson: Garbage in = garbage out.

[Vincent @joqivi]

Auto focus, grain and lens distortion in the browser with a 10 million #splat scene looks soo good! Setup:
#Osmo360
6gb 8k scene
800 sampled views per lens.
10m point cloud
Trained from raw osv file in the browser using precession.

[Vincent @joqivi]

Been learning how to make a #splat from my #osmo360 camara. The main challenge has been how to utilise the most amount of the lens data for pose estimation and training while not poisoning the pipeline with seems, moving elements and lens distortion.

[Vincent @joqivi]

Or so I thought.. My computer cant even run Gemma4 lol. So new computer or pay the overlords an ever increasing intelligence utility tax forever. Local for sure. Wonder what machine I should buy. Suggestions?

[Vincent @joqivi]

@claudeai #fable5 is dope. Ran out of credits in like 2s...
Switch to @opencode for 3h lol.

1. Free models train on your code...
2. xai build 0.0.1 can't even center a div 3.
@ollama + @opencode #gemma4 is promising.

Anyone else apocalypse and AI overlord proofing?

[Vincent @joqivi]

@DuckbillStudio Rigging splats is fucking cool. How does it know where to put the bones??

[Vincent @joqivi]

@IlirAliu_ This is sick! How does it handle longer sequences? Drift?

[Vincent @joqivi]

Conclusion: sometimes simple is better. Remove parts, don't add complexity for the sake of complexity.

[Vincent @joqivi]

360 capture, what I learned. TLDR:
- Forget iPhone capture, 360 is the way.
- SAM 3 to mask out the operator and seam.
- Train on raw lenses, no baked seam.
- Don't reproject to pinhole, use the raw lenses.
- COLMAP: use_ut on. gsplat: MCMC is the only supported strategy.

[Vincent @joqivi]

Honest part: PSNR 19 on fisheye-only, lower than it should be, probably a gsplat config knob not the capture. What trainer are you running on 360 footage? Curious how others handle the seam and the operator.

[Vincent @joqivi]

6/ One thing's certain: 360 still needs a lot of views to look good. The cost math is brutal. But if I can crack the 360-native path instead of brute-forcing pinhole reprojection, it would be so sick. More next week.

[Vincent @joqivi]

1/ I'm teaching myself 3D reconstruction properly. I want to build as much of the pipeline myself as I can, so I understand it in depth instead of gluing black boxes together. This week's test: reconstruct a single tree with my new Osmo 360. Sounds trivial. It's not.

[Vincent @joqivi]

5/ So I'm weighing going back to a 360-only pipeline for point cloud generation. Maybe revisiting the full thing as 2 separate fisheye lenses, now that I actually understand how to set the rig up in COLMAP.

[Vincent @joqivi]

Okay Radial Camera Model was the solve. Points are dead center now. A bit of a backwards approach since I know there are splat segmentation models these days but I think its important to start simple when doing something complicated lol. Totally works if the point cloud is solid.

[Vincent @joqivi]

Experimenting with spatial indexing. Was using an VLM to tag up an image. COLMAP+GLOMAP then raycast through the 2d coords into the 3d scene to create a 3d tag. Works well on the y axis by not the x axis. Turns out OPENCV do not make square pixels intrinsics. Radial might work.