What is anthrotap?

cvlab-kaist/anthrotap — explained in plain English

Analysis updated 2026-07-15 · repo last pushed 2026-06-23

★ 32PythonAudience · researcherComplexity · 4/5ActiveSetup · hard

Mindmap

mindmap
  root((repo))
    What it does
      Tracks body points in video
      Fine-tunes LocoTrack model
      Uses real human motion data
    Tech stack
      Python
      LocoTrack
      Kubric synthetic data
    Use cases
      Motion capture without suits
      Sports technique analysis
      Augmented reality effects
    Audience
      Computer vision researchers
      AR developers
      Sports analysts
    Outputs
      Pre-trained model weights
      Interactive local demo
      Benchmark evaluation scripts

mindmap root((repo)) What it does Tracks body points in video Fine-tunes LocoTrack model Uses real human motion data Tech stack Python LocoTrack Kubric synthetic data Use cases Motion capture without suits Sports technique analysis Augmented reality effects Audience Computer vision researchers AR developers Sports analysts Outputs Pre-trained model weights Interactive local demo Benchmark evaluation scripts

Click or tap to explore — scroll the page freely

What do people build with it?

USE CASE 1

Build markerless motion capture by tracking body points from regular video footage.

USE CASE 2

Analyze sports technique by tracking specific joints or body parts through athletic movements.

USE CASE 3

Create augmented reality effects that anchor visual elements to a person's body as they move.

USE CASE 4

Evaluate point-tracking accuracy on standard human motion benchmark datasets.

What is it built with?

PythonPyTorchLocoTrackKubric

How does it compare?

	cvlab-kaist/anthrotap	autolearnmem/automem	cortex-ai-network/crypto-arbitrage-bot-automated-trading
Stars	32	32	32
Language	Python	Python	Python
Last pushed	2026-06-23	—	—
Maintenance	Active	—	—
Setup difficulty	hard	hard	moderate
Complexity	4/5	5/5	2/5
Audience	researcher	researcher	general

Figures from each repo's GitHub metadata at analysis time.

How do you get it running?

Difficulty · hard Time to first run · 1h+

Requires PyTorch with CUDA GPU, downloading large pre-trained model weights, and configuring benchmark datasets plus environment setup for the LocoTrack model.

No license information is provided in the repo, so you would need to contact the authors for permission before using it.

So what is it?

AnthroTAP is a research project that helps computers track specific points on a person's body as they move through a video. Imagine picking a spot on someone's hand in one frame and having the system automatically follow that exact spot as the person waves, dances, or walks around, that's what this improves. The problem it solves is that existing point-tracking systems are mostly trained on artificial, computer-generated scenes, which don't capture how real humans move. People bend, twist, and articulate in complex ways that synthetic data doesn't reproduce well. AnthroTAP fine-tunes an existing tracker called LocoTrack using a curated dataset of real-world human motion videos with dense point annotations, so the tracker gets much better at following points on people in actual footage. Researchers in computer vision would use this, for example, someone building motion capture without special suits, analyzing sports technique from regular video, or developing augmented reality effects that need to stick to a person's body. The project includes pre-trained model weights you can download and evaluate on standard benchmark datasets, plus an interactive demo you can run locally to try it on your own videos. The training pipeline combines synthetic data (Kubric) with the real human motion data, and the project provides configuration files and scripts to manage both. The README doesn't go into detail on exactly how the human motion dataset was created or annotated. The project comes from a collaboration between KAIST, Adobe Research, the University of Michigan, and LG AI Research, and is associated with a CVPR 2026 paper.

Copy-paste prompts

Prompt 1

I want to use AnthroTAP to track points on a person in my video. Walk me through downloading the pre-trained model weights, setting up the environment, and running the interactive demo locally on an MP4 file.

Prompt 2

Help me set up the AnthroTAP evaluation pipeline on a standard benchmark dataset. What configuration files do I need to modify, and how do I run the evaluation scripts to get tracking accuracy metrics?

Prompt 3

I want to understand how AnthroTAP fine-tunes LocoTrack with real human motion data. Explain the training pipeline that combines Kubric synthetic data with the real human motion dataset, and help me configure the training scripts.

Prompt 4

I have the AnthroTAP repo cloned and PyTorch installed. What dependencies are missing, what model weights do I need to download, and how do I run point tracking on my own video file step by step?

Frequently asked questions