TencentARC

TencentARC / Track4World

Public

Track4World: Feedforward World-centric Dense 3D Tracking of All Pixels

3d 3dreconstruction 3dtracking depthanything moge-2
19
0
100% credibility
Found Mar 04, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Track4World enables dense 3D tracking of every pixel in monocular videos using world-centric coordinates for accurate scene flow estimation.

How It Works

1
🔍 Discover Track4World

You stumble upon this exciting tool that promises to track every single pixel in 3D from everyday videos.

2
📹 Grab your video

Pick a fun clip like a cat playing or cars zooming by to see motion come alive.

3
⚙️ Quick setup

Follow easy steps to prepare everything, including grabbing ready-to-use brainpower files.

4
🚀 Launch the magic

Click run and watch it analyze your video, revealing hidden 3D movements instantly.

5
Choose your view
📏
Flat 2D paths

See colorful lines tracing motion across the screen.

🌍
Full 3D world

Dive into a spinning 3D scene with every pixel's journey.

6
👀 Explore results

Zoom, rotate, and play with glowing trails and point clouds showing real motion.

✨ Your 3D motion unlocked

You now understand exactly how everything moves in true 3D space, ready for videos, analysis, or fun!

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is Track4World?

Track4World is a Python toolkit for dense, feedforward 3D tracking of all pixels in monocular videos, outputting world-centric scene flow and trajectories without iterative refinement. It solves the challenge of lifting 2D video motion into consistent 3D world coordinates, enabling reconstruction of dynamic scenes from raw footage. Users get pre-trained models on HuggingFace, CLI demos for 2D/3D tracking (camera- or world-centric), and eval scripts for benchmarks like KITTI or TAP-Vid.

Why is it gaining traction?

It stands out by delivering pixel-dense tracking in a single feedforward pass—fast and global, unlike iterative or sparse alternatives. Developers hook into ready weights for Depth Anything or Pi3 depth, plus SAM2 segmentation for dynamic objects, yielding PLY point clouds and 4D viz gifs out-of-box. Early adopters praise the no-fuss setup for monocular 3D motion on arbitrary videos.

Who should use this?

Robotics engineers tracking objects in real-time egocentric video, AR devs reconstructing dynamic environments from phone cams, or CV researchers prototyping dense scene flow on custom datasets. Ideal for those ditching sparse trackers like TAP-Vid for full-pixel density without multi-view setups.

Verdict

Grab it for quick monocular 3D prototypes—demos run smoothly on CUDA 12/Python 3.11—but at 19 stars and 1.0% credibility, it's raw; expect tweaks for production. Solid README and evals make it worth forking now.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.