---

🚀 News

• (2025.12.28) 🔥 We have updated the Training and Evaluation code! Check out the new sections below.

Overview

TAPIP3D is a method for long-term feed-forward 3D point tracking in monocular RGB and RGB-D video sequences. It introduces a 3D feature cloud representation that lifts image features into a persistent world coordinate space, canceling out camera motion and enabling accurate trajectory estimation across frames.

We provide a detailed video illustration of our TAPIP3D.

Installation

Installing dependencies

• Prepare the environment

conda create -n tapip3d python=3.10
conda activate tapip3d
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 "xformers>=0.0.27" --index-url https://download.pytorch.org/whl/cu124
pip install torch-scatter -f https://data.pyg.org/whl/torch-2.4.1+cu124.html
pip install -r requirements.txt

• Compile pointops2

cd third_party/pointops2
LIBRARY_PATH=$CONDA_PREFIX/lib:$LIBRARY_PATH python setup.py install
cd ../..

• Compile megasam

cd third_party/megasam/base
LIBRARY_PATH=$CONDA_PREFIX/lib:$LIBRARY_PATH python setup.py install
cd ../../..

Downloading checkpoints

Download our TAPIP3D model checkpoint here to checkpoints/tapip3d_final.pth

If you want to run TAPIP3D on monocular videos, you need to prepare the following checkpoints manually to run MegaSAM:

• Download the DepthAnything V1 checkpoint from here and put it to third_party/megasam/Depth-Anything/checkpoints/depth_anything_vitl14.pth
• Download the RAFT checkpoint from here and put it to third_party/megasam/cvd_opt/raft-things.pth

Additionally, the checkpoints of MoGe and UniDepth will be downloaded automatically when running the demo. Please make sure your network connection is available.

Demo Usage

We provide a simple demo script inference.py, along with sample input data located in the demo_inputs/ directory.

The script accepts as input either an .mp4 video file or an .npz file. If providing an .npz file, it should follow the following format:

• video: array of shape (T, H, W, 3), dtype: uint8
• depths (optional): array of shape (T, H, W), dtype: float32
• intrinsics (optional): array of shape (T, 3, 3), dtype: float32
• extrinsics (optional): array of shape (T, 4, 4), dtype: float32

For demonstration purposes, the script uses a 32x32 grid of points at the first frame as queries.

Inference with Monocular Video

By providing a video as --input_path, the script first runs MegaSAM with MoGe to estimate depth maps and camera parameters. Subsequently, the model will process these inputs within the global frame.

Demo 1

To run inference:

python inference.py --input_path demo_inputs/sheep.mp4 --checkpoint checkpoints/tapip3d_final.pth --resolution_factor 2

An npz file will be saved to outputs/inference/. To visualize the results:

python visualize.py 

Demo 2

python inference.py --input_path demo_inputs/pstudio.mp4 --checkpoint checkpoints/tapip3d_final.pth --resolution_factor 2

Inference with Known Depths and Camera Parameters

If an .npz file containing all four keys (rgb, depths, intrinsics, extrinsics) is provided, the model will operate in an aligned global frame, generating point trajectories in world coordinates. We provide one example .npz file at here and please put it in the demo_inputs/ directory.

Demo 3

python inference.py --input_path demo_inputs/dexycb.npz --checkpoint checkpoints/tapip3d_final.pth --resolution_factor 2

Training and Evaluation

1. Dataset Preparation

Please refer to DATASET.md for instructions on preparing datasets for both training and evaluation.

2. Training

To start training, run:

bash scripts/train.sh

• experiment_name: The run name shown on WandB.
• experiment_id: A unique identifier. Re-running with the same experiment_id will automatically resume training from the latest checkpoint.

3. Evaluation

To evaluate a checkpoint, run:

bash scripts/eval.sh

You can specify the model to evaluate by modifying the checkpoint variable in scripts/eval.sh.

Citation

If you find this project useful, please consider citing:

@article{tapip3d,
  title={TAPIP3D: Tracking Any Point in Persistent 3D Geometry},
  author={Zhang, Bowei and Ke, Lei and Harley, Adam W and Fragkiadaki, Katerina},
  journal={arXiv preprint arXiv:2504.14717},
  year={2025}
}

--- Tranlated By Open Ai Tx | Last indexed: 2026-02-12 ---