Effxcts

🚨 Just built a system that reconstructs 4D NBA plays from a single broadcast video. 📦 Pipeline includes: • Object Detection • Camera Stabilization • Pose Tracking • Homography • Ball Path See more examples on my page—happy to answer any questions below.

the one singular ball in vegas is so funny to me for no reason

Longevity.

83

@yanpei_cao Very cool. Does it work with photorealistic humans?

3D generation often treats objects as single, monolithic shapes. Our new research, OmniPart, takes a different path: generating objects as assemblies of parts for structural coherence and fine-grained control. And it has been accepted to SIGGRAPH Asia 2025 🎉

@1littlecoder check my pinned post for 3D pose reconstruction of nba plays

The coolest thing you'll see on computer vision today!! Incredible accuracy and great attention to detail.

@ChrisAtKIRI Does it work well on a headshot?

Happy to have a chance to test the latest Rodin Gen2 by Deemos! The result speaks for itself! Haven't had a chance to try "Bang" generative parts yet, but Gen2 is truly impressive! 1.5 to 2 is a huge improvement, I can tell! #3D #GenerativeAI

@skalskip92 @903124S example of stabilized clip with poses

@EffxctsLLC @903124S that is such a headache… this model is still the weakest part of that pipeline. you can see how those virtual markers wiggle… that’s because the model is still not accurate enough.

That's what would happen if you don't map 2d pose estimation to 3d properly (I still have no idea how to account for 3d position of players if they're jumping) @skalskip92

@skalskip92 @903124S yes I figured it would be tough to have an accurate model for court key point detection. stabilizing the camera in the pipeline with third-party software like After Effects makes it a lot easier. there might be other options for closer to real-time stabilization

@903124S @skalskip92 yes i detect jumps and then interpolate from beginning of jump position to end of jump position in the homography. and then i use the time of the jump to determine z height

@EffxctsLLC @skalskip92 As I understand it works in a way that when you detect jump it transform homography differently base on if they're jumping?

@skalskip92 @903124S yes this was before your nba keypoint detection model was out and I didn’t want to create a dataset to train a model lol

@EffxctsLLC @903124S camera stabilization as part of the pipeline? that’s a really interesting idea

@ODJ__ vitpose+-h

@EffxctsLLC What are you using for pose estimation

real hoopers know whose POV this is

@CrumpledJumper 😂

@EffxctsLLC I'm not saying

All of Jokic’s career FG visualized

All of Tatum’s career FG visualized

@SkyedOKC More 3d data visualizations on the way 👈

all of SGAs career FGs (via @EffxctsLLC)

All of SGA’s career FG visualized