SpatialTrackerV2: 3D Point Tracking Made Easy

CAD&CG, Zhejiang University; University of Oxford; Ant Research; Pixelwise AI; NUS

Yuxi Xiao, Jianyuan Wang, Nan Xue, Nikita Karaev, Iurii Makarov, Bingyi Kang, Xin Zhu, Hujun Bao, Yujun Shen, Xiaowei Zhou

Project Page | Paper V1 | Paper V2 | BibTeX | Goolge Drive

TODO List

Release Stage 1 Checkpoint
- SpaTrack2+ MegaSAM ➔ supports Posed RGBD as input.
- SpaTrack2 + Moge ➔ supports unposed RGB as input.
Release Stage 2 Checkpoint
- Track-guided Depth Stablizer.
Release Stage 3 Checkpoint
- Sliding windows version.
More Releases
- Some Ceres Python Bindings designed for SpatialTracker and Dynamic Reconstruction.
- More supports for other Depth Model, i.e., DepthAnything, StereoFoundation, UniDepth, Metric3D.

Set up the environment

To set up the environment for running the SpaTrack model, follow these steps:

Clone the Repository:

git clone git@github.com:henry123-boy/SpaTrack2.git
cd SpaTrack2

Create a Virtual Environment: It's recommended to use a virtual environment to manage dependencies.
```
conda create -n SpaTrack2 python=3.11
conda activate SpaTrack2
```

Install Dependencies:

Install the torch dependencies pip (tested with torch2.4).

python -m pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu124

Install the required Python packages using pip.

python -m pip install -r requirements.txt

Install SpaTrack2 Visualizer:
```
cd viser
python -m pip install -e .
```

By following these steps, you should have a working environment ready to run the SpaTrack model.

Download the Checkpoints

mkdir -p checkpoints

Step1: Download the checkpoint of Moge from here, and put the model.pt into ./checkpoints/

Step2: Download the checkpoint of SpaTrack2 from GoolgeDrive, and place it into ./checkpoints/

Quick Start

We gave two examples to illustrate the usage of SpaTrack2. Firstly, please download ckpts and examples via:

sh scripts/download.sh

Type1: Posed RGBD video (Example0)

We provide an example who has Posed RGBD input with MegaSAM.

python inference.py --data_type="RGBD" --data_dir="assets/example0" --video_name="snowboard" --fps=1

Type2: unposed RGB video (Example1)

python inference.py --data_type="RGB" --data_dir="assets/example1" --video_name="xhs" --fps=6

Visualize your results

We provide two types of visualization. The guidance will be displayed in the terminal after running inference.py

🌟 Recommended: Gradio Demo with SAM 🌟

Please follow the instructions in the app_3rd README to configure the dependencies. Then,

python -m pip install gradio==5.31.0 pako

Our gradio demo enable the user to track the points on the target object easily, just try:

python app.py

Note❗

For users employing vscode as their IDE, the 3D trajectory visualization will automatically appear. Ensure that the port is accessible and consistent both locally and on the server. If not, you may need to manually forward the port using SSH -L.

4.5 KiB Raw Blame History