Point Tracking for 4DGS

This repo is the point tracking Python extension for 4D Gaussian Splatting. It wraps sequence point trackers such as CoTracker3, VGGT, MV-TAP, OpenD4RT, MATCHA, MV-RoMa, and UniFlowMatch behind one small registry, then applies them either to plain image sequences or to multi-timestep Gaussian Splatting camera datasets.

The package provides three common workflows:

• track sampled 2D points through one ordered image sequence
• project 3D Gaussians into a reference timestep, track those projected points across all timesteps, and attach the resulting tracks back to Gaussian Splatting camera datasets
• build supervision samples (query, frames, and per-view tracks) from camera datasets, a trained 3DGS point cloud, or a Blender scene, and save them for later training

Features

• Organised as a standard Python package with pip install support
• Shared point tracker registry with cotracker3, vggt, vggsfm, mvtap, d4rt, matcha, mvroma, uniflowmatch, and blender
• Single-view image sequence tracking and rendering
• Multi-timestep Gaussian Splatting camera dataset tracking
• Point batching through --batch-size for long tracks
• Camera dataset reordering against a selected reference timestep
• Blender rigid-mesh tracking for scenes rendered by NVS2COLMAP
• Supervision datasets (tracker, tracker_gs, blender) that save query, frames, and tracks
• Port the motion estimation workflow from TrackerSplat for point-tracker-driven motion synthesis
• Regularize 3DGS training with point tracker trajectories

Install

Prerequisites

• PyTorch (CUDA build recommended)
• CUDA Toolkit matching your PyTorch installation
• gaussian-splatting >= 2.3.8

pip install track-4dgs also installs Pillow, einops, timm, PyYAML, numpy, opencv-python, warp-lang, and diffusers>=0.30,<0.36. Keep diffusers below 0.36 so it does not pull huggingface-hub 1.x, which breaks transformers<5.

The published wheel leaves out Git dependencies. Install them for the trackers that need them:

pip install wheel setuptools
pip install --upgrade "gaussian-splatting>=2.3.8"

# cotracker3
pip install --upgrade git+https://github.com/facebookresearch/co-tracker.git@main

# vggt and vggsfm
pip install --upgrade hydra-core omegaconf
pip install --upgrade git+https://github.com/facebookresearch/vggt.git@main
pip install --upgrade "git+https://github.com/jytime/LightGlue.git#egg=lightglue"

# uniflowmatch and mvroma
pip install --upgrade "git+https://github.com/castacks/UniCeption.git@ee7fa0b50d8295b3f0485e42c1d133fb534bfce7"

matcha, mvroma, and uniflowmatch ship inside this package. Their source trees are the submodules/matcha, submodules/MV-RoMa, and submodules/UFM submodules, so a source install needs git clone --recursive.

If you have trouble with gaussian-splatting, try installing it from source:

pip install wheel setuptools
pip install --upgrade git+https://github.com/yindaheng98/gaussian-splatting.git@master --no-build-isolation

PyPI Install

pip install --upgrade track-4dgs

or build latest from source:

pip install wheel setuptools
pip install --upgrade git+https://github.com/yindaheng98/track-4dgs.git@master --no-build-isolation

Development Install

git clone --recursive https://github.com/yindaheng98/track-4dgs.git
cd track-4dgs
pip install --target . --upgrade . --no-deps

--no-deps skips everything in Prerequisites. Install those packages first.

Download Checkpoints

CoTracker3 offline checkpoint (cotracker3, default checkpoints/scaled_offline.pth):

mkdir -p checkpoints
wget -P checkpoints https://huggingface.co/facebook/cotracker3/resolve/main/scaled_offline.pth

VGGT commercial checkpoint (vggt, default checkpoints/vggt_1B_commercial.pt):

mkdir -p checkpoints
wget -P checkpoints https://huggingface.co/facebook/VGGT-1B-Commercial/resolve/main/vggt_1B_commercial.pt --header="Authorization: Bearer $HF_TOKEN"

If the VGGT checkpoint is not present, VGGTPointTracker falls back to VGGT.from_pretrained("facebook/VGGT-1B"). VGGT reads query features from frame 0, so every query frame index must be 0.

VGGSfM tracker checkpoint (vggsfm, default checkpoints/vggsfm_v2_tracker.pt):

mkdir -p checkpoints
wget -P checkpoints https://huggingface.co/facebook/VGGSfM/resolve/main/vggsfm_v2_tracker.pt

If the checkpoint is not present, load_vggsfm downloads it from that URL.

MATCHA checkpoint (matcha, default checkpoints/matcha_pretrained.pth, from the MATCHA release):

mkdir -p checkpoints
# save the downloaded file as checkpoints/matcha_pretrained.pth

MV-TAP checkpoint (mvtap, default checkpoints/mvtap.ckpt, from the MV-TAP release):

mkdir -p checkpoints
# save the downloaded file as checkpoints/mvtap.ckpt

MV-RoMa checkpoint (mvroma, default checkpoints/outdoor_final.pth, from the MV-RoMa release). Indoor weights are here. The tracker also downloads infinity1096/UFM-Refine as its prematcher. It runs on CUDA and accepts at most 31 reference frames.

mkdir -p checkpoints
# save the downloaded file as checkpoints/outdoor_final.pth

UniFlowMatch (uniflowmatch) downloads a Hugging Face repo on first use. -m model="'base'" selects infinity1096/UFM-Base. Other names are refine, base-980, refine-980, base-dinov2l-init, and base-dinov2g-init.

OpenD4RT checkpoint and model config (d4rt, default checkpoints/OpenD4RT_48CLIP_9Mix_NoCropAUG/opend4rt.ckpt):

# 48-frame model
mkdir -p checkpoints/OpenD4RT_48CLIP_9Mix_NoCropAUG
wget -P checkpoints/OpenD4RT_48CLIP_9Mix_NoCropAUG https://huggingface.co/Lijiaxin0111/OpenD4RT/resolve/main/checkpoints/OpenD4RT_48CLIP_9Mix_NoCropAUG/opend4rt.ckpt
wget -P checkpoints/OpenD4RT_48CLIP_9Mix_NoCropAUG https://huggingface.co/Lijiaxin0111/OpenD4RT/resolve/main/checkpoints/OpenD4RT_48CLIP_9Mix_NoCropAUG/model.yaml

# 32-frame model
mkdir -p checkpoints/OpenD4RT_32CLIP_9Dataset_NoAUG
wget -P checkpoints/OpenD4RT_32CLIP_9Dataset_NoAUG https://huggingface.co/Lijiaxin0111/OpenD4RT/resolve/main/checkpoints/OpenD4RT_32CLIP_9Dataset_NoAUG/opend4rt.ckpt
wget -P checkpoints/OpenD4RT_32CLIP_9Dataset_NoAUG https://huggingface.co/Lijiaxin0111/OpenD4RT/resolve/main/checkpoints/OpenD4RT_32CLIP_9Dataset_NoAUG/model.yaml

Each model requires both opend4rt.ckpt and its adjacent model.yaml. Pass either the checkpoint file or its containing directory through the checkpoint tracker option.

Command-Line Usage

List Registered Point Trackers

Verify that track_4dgs can import and register trackers:

python -c "import track_4dgs; print(track_4dgs.get_available_point_trackers())"

track1v accepts an AbstractViewPointTracker. Joint trackers run through build_dataset. blender is documented in Blender mesh tracking.

Track One Image Sequence

python -m track_4dgs.track1v \
    -s data/frame_000.png data/frame_001.png data/frame_002.png \
    -d output/track1v-cotracker3 \
    --tracker cotracker3 \
    --num-points 256

The command samples random query points on the first image, tracks them through all source images, and saves a rendered track overlay sequence under the destination directory. Overlay color follows confidence: red is low, green is high.

Tracker-specific options can be passed with repeated -m/--option_tracker values. --batch-size forwards chunks of query points to the tracker; omit it to track every point in one call.

python -m track_4dgs.track1v \
    -s data/frame_000.png data/frame_001.png data/frame_002.png \
    -d output/track1v-vggt \
    --tracker vggt \
    --batch-size 128 \
    -m checkpoint="'checkpoints/vggt_1B_commercial.pt'" \
    -m iters=4

Build a Supervision Dataset

python -m track_4dgs.build_dataset runs a registered supervision dataset. --save-folder and --vis-folder are both optional, and at least one is required. With --save-folder, it writes one folder per sample:

<save-folder>/supervision<index>/
    query.pt
    frame1/ ...          # COLMAP images and sparse/0
    0.pt 1.pt ...        # one Track per view

--save-start is the first index and defaults to 1, so the first saved sample is supervision1. --vis-folder writes rainbow overlays. If --save-folder is set, the overlays are drawn from the saved samples. If it is omitted, build_dataset draws the in-memory dataset directly and does not write it to disk. Omit --vis-folder to save without overlays.

List the dataset names:

python -c "import track_4dgs; from track_4dgs.dataset.registry import get_available_supervision_datasets; print(get_available_supervision_datasets())"

• tracker: sample points uniformly inside the scene cube, keep those seen by more cameras than --visible-cameras (default 2), project them, and track
• tracker_gs: sample a random subset of a trained 3DGS point cloud, then track those points
• blender: render a .blend file from sampled cameras and track the query on its rigid meshes. See Blender mesh tracking

Uniform scene points:

python -m track_4dgs.build_dataset \
    --dataset tracker \
    --tracker cotracker3 \
    -s data/sequence/frame_000 data/sequence/frame_001 data/sequence/frame_002 \
    --num-points 256 \
    --visible-cameras 2 \
    --save-folder output/sequence/dataset-cotracker3 \
    --vis-folder output/sequence/dataset-cotracker3-vis

Gaussian point cloud. -d and -i name the run whose checkpoint is sampled:

<destination>/point_cloud/iteration_<iteration>/point_cloud.ply

python -m track_4dgs.build_dataset \
    --dataset tracker_gs \
    --tracker mvtap \
    -s data/sequence/frame_000 data/sequence/frame_001 data/sequence/frame_002 \
    -d output/sequence \
    -i 30000 \
    --num-points 256 \
    --batch-size 32 \
    --save-folder output/sequence/dataset-mvtap \
    --vis-folder output/sequence/dataset-mvtap-vis

Repeat -s/--sources to add another sample. --load_cameras and --init-dataset-index attach to the preceding --sources and may each appear once in that group. --init-dataset-index defaults to 0. For tracker_gs, repeat -d and -i once per --sources.

Blender Mesh Tracking

track_4dgs.blender (--tracker blender) binds query pixels to rigid meshes in a .blend file and carries them with each mesh matrix_world. Use it together with the Blender tools in NVS2COLMAP.

NVS2COLMAP renders that blend file into a Neural 3D Video scene (poses_bounds.npy and one video per camera), then nvs2colmap.n3dv writes one COLMAP camera dataset per frame. This tracker reads the same .blend file and those frame datasets. Cameras stay in the blend world, look along +Z, and treat image +Y as downward, which is the OpenCV camera NVS2COLMAP stores in poses_bounds.npy. Dataset frame i in --sources is Blender frame frame_start + i * frame_step, and that frame is at most frame_end. Pass the same --frame_start, --frame_end, and --frame_step as nvs2colmap.blender. frame_start defaults to 1. frame_end and frame_step default to the blend scene. Install NVS2COLMAP and keep a Blender executable on PATH as blender, or pass another path with -m blender="'path/to/blender'". warp-lang is already installed with this package.

nvs2colmap-sample-cameras can place the cameras and write poses_bounds.npy. Render the blend file from those cameras:

python -m nvs2colmap.blender \
    --path data/scene \
    --blend data/scene.blend

Decode the rendered videos into per-frame COLMAP datasets:

python -m nvs2colmap.n3dv \
    --path data/scene \
    --ffmpeg ffmpeg \
    --ffprobe ffprobe

Track those frames with the same blend file. This only writes overlays:

python -m track_4dgs.build_dataset \
    --dataset tracker_gs \
    --tracker blender \
    -s data/scene/frame1 data/scene/frame2 data/scene/frame3 \
    -d output/scene \
    -i 30000 \
    -m blend="'data/scene.blend'" \
    --vis-folder output/scene/track-blender-vis

Copy --frame_start, --frame_end, and --frame_step from the render command onto the tracker. nvs2colmap.blender --frame_step already drops frames inside each camera video, and nvs2colmap.n3dv still writes consecutive frame* folders. Leave nvs2colmap.n3dv --start-number at its default of 1: that flag counts frames inside the rendered video, so frame1 is the first frame the Blender render wrote. --sources then lists frame1, frame2, ... in order.

python -m nvs2colmap.blender \
    --path data/scene \
    --blend data/scene.blend \
    --frame_start 10 \
    --frame_end 14 \
    --frame_step 2

python -m track_4dgs.build_dataset \
    --dataset tracker_gs \
    --tracker blender \
    -s data/scene/frame1 data/scene/frame2 data/scene/frame3 \
    -d output/scene \
    -i 30000 \
    -m blend="'data/scene.blend'" \
    -m frame_start=10 \
    -m frame_end=14 \
    -m frame_step=2 \
    --vis-folder output/scene/track-blender-vis

build_dataset --dataset blender can skip the COLMAP export and render supervision directly. It samples OpenCV cameras on a spherical cap, samples the query near the mesh at --init-dataset-index, renders every view, and binds the tracks to the rigid meshes. --frame_start, --frame_end, --frame_step, and --init-dataset-index attach to the preceding --blend. Repeat --blend for another sample.

python -m track_4dgs.build_dataset \
    --dataset blender \
    --blend data/scene.blend \
    --n-cameras 8 \
    --height 512 \
    --width 512 \
    --focal 500 \
    --num-points 256 \
    --save-folder output/scene/dataset-blender \
    --vis-folder output/scene/dataset-blender-vis

API Usage

Build a Tracker

from track_4dgs.registry import build_point_tracker, get_available_point_trackers

print(get_available_point_trackers())

tracker = build_point_tracker(
    "cotracker3",
    checkpoint="checkpoints/scaled_offline.pth",
).to("cuda")

Track Image Tensors

sample_query returns pixel coordinates and a frame index of 0 for every point. track_view is the single-view entry used by track1v.

import torch

from track_4dgs.track1v import load_image, sample_query

frames = [
    load_image("data/frame_000.png", "cuda"),
    load_image("data/frame_001.png", "cuda"),
]
points, frame_indices = sample_query(frames[0], num_points=256)

with torch.no_grad():
    track = tracker.track_view(
        points,
        frame_indices,
        frames,
        [None] * len(frames),
    )

print(track.points.shape)      # [num_frames, num_points, 2]
print(track.visibility.shape)  # [num_frames, num_points]
print(track.confidence.shape)  # [num_frames, num_points]

Track Gaussian Splatting Camera Datasets

build_camera_dataset_tracker and CameraDatasetTracker are reserved. They attach each view's track onto the cameras as camera.custom_data["track"]. build_dataset does not call them; it keeps the view-major Supervision tracks.

from gaussian_splatting.prepare import prepare_gaussians

from track_4dgs.dataset import build_camera_dataset_tracker, prepare_datasets, query_views_from_gaussians

sources = [
    "data/sequence/frame_000",
    "data/sequence/frame_001",
]
datasets = prepare_datasets(
    sources=sources,
    device="cuda",
    reorder_reference_idx=0,
)
gaussians = prepare_gaussians(
    sh_degree=3,
    source=sources[0],
    device="cuda",
    load_ply="output/sequence/point_cloud/iteration_30000/point_cloud.ply",
)
dataset_tracker = build_camera_dataset_tracker(
    tracker_name="cotracker3",
    device="cuda",
    tracker_configs={"checkpoint": "checkpoints/scaled_offline.pth"},
)

queries = query_views_from_gaussians(
    datasets=datasets,
    gaussians=gaussians,
    init_dataset_index=0,
    num_points=256,
)
tracked_datasets = dataset_tracker(queries, datasets, batch_size=128)

camera_track = tracked_datasets[0][0].custom_data["track"]
print(camera_track.points.shape)  # [num_points, 2] for this camera in this frame

Load Saved Supervision

from track_4dgs import SavedDataset

dataset = SavedDataset("output/sequence/dataset-mvtap", start=1)
sample = dataset[0]

print(sample.query.points.shape)       # [num_views, num_points, 2]
print(len(sample.frames))               # num_frames
print(sample.tracks[0].points.shape)    # [num_frames, num_points, 2]

frame = sample.tracked_camera_dataset(0)
camera_track = frame[0].custom_data["track"]
print(camera_track.points.shape)        # [num_points, 2] for this camera in this frame

Supervision holds a view-major Query, one camera dataset per frame, and one Track per view. AbstractDataset.dump writes that layout; SavedDataset reads it back. tracked_camera_dataset(index) returns that frame as a TrackedCameraDataset: each camera's custom_data["track"] is the CameraTrack sliced from the matching view track. This does not run a tracker; the tracks are already stored on the sample.

Design: Point Tracker Registry

AbstractPointTracker.__call__ validates inputs, then calls track. It takes a view-major Query (points [V, N, 2], frame_indices [V, N], in_image [V, N]) and frame-major camera datasets, and returns one Track per view. batch_size splits the point dimension; None tracks every point in one call. build_dataset keeps that view-major Supervision layout.

CameraDatasetTracker and build_camera_dataset_tracker are reserved for a later step that hangs those tracks back on the cameras. build_camera_dataset_tracker(tracker_name, device, tracker_configs) builds a registered point tracker and wraps it. Calling the wrapper with a view-major Query and frame-major camera datasets returns one TrackedCameraDataset per frame. Each camera then carries its CameraTrack on camera.custom_data["track"]. A saved sample can rebuild one frame of that layout with Supervision.tracked_camera_dataset(index) without running the tracker again. build_dataset does not call this path:

query + frames -> validate + track -> view_tracks -> tracked camera datasets

Single-view trackers (AbstractViewPointTracker, e.g. CoTracker3 / VGGT / MATCHA) split the query by view and implement track_batch(points, frame_indices, frames, masks) for one camera sequence. Joint trackers (AbstractBatchPointTracker, e.g. MV-TAP / Blender) consume all views at once through track_batch(query, frames) and batch along the point dimension.

Frame 0 cameras --\
Frame 1 cameras ----> single-view loop, or one joint multi-view call ----> view_tracks
Frame 2 cameras --/

in_image marks query pixels that project inside the frame (and in front of the camera, when the query comes from Query.from_projection). Single-view trackers skip points that fail this test. Joint trackers fold it into Track.mask.

This keeps model-specific code isolated in tracker implementations while the 4DGS workflow can use any registered tracker by name.

Extending: Adding a New Point Tracker

Single-view trackers return Track(points=[D, N, 2], visibility=[D, N], confidence=[D, N], mask=[D, N]) from track_batch:

from collections.abc import Sequence

import torch

from track_4dgs.tracker import AbstractViewPointTracker, Track


class MyPointTracker(AbstractViewPointTracker):
    def to(self, device):
        self.device = torch.device(device)
        return self

    def track_batch(
        self,
        points: torch.Tensor,
        frame_indices: torch.Tensor,
        frames: Sequence[torch.Tensor],
        frame_masks: Sequence[torch.Tensor | None],
    ) -> Track:
        ...

A joint tracker subclasses AbstractBatchPointTracker and implements track_batch(self, query, frames) -> Sequence[Track], one Track per view.

Register it at import time:

from track_4dgs.registry import register_point_tracker

register_point_tracker("mytracker", MyPointTracker)

After registration, it is available everywhere a single-view tracker is accepted:

python -m track_4dgs.track1v --tracker mytracker -s frame0.png frame1.png -d output/mytracker

Import the module that calls register_point_tracker before build_point_tracker. The built-in trackers register themselves from track_4dgs/__init__.py.

Acknowledgement

This repo is developed based on CoTracker, VGGT, MV-TAP, OpenD4RT, MATCHA, MV-RoMa, UniFlowMatch, UniCeption, LightGlue, gaussian-splatting (packaged), and NVS2COLMAP. Many thanks to the authors for open-sourcing their codebases.