MVOT: MatrixCity Video Object Tracking

MVOT is a synthetic benchmark for video object tracking in satellite imagery. It contains short urban-scene videos rendered from MatrixCity, with frame-level target bounding boxes across controlled illumination, camera-tilt, trajectory, and occlusion conditions.

The dataset accompanies SatSAM2: Motion-Constrained Video Object Tracking in Satellite Imagery using Promptable SAM2 and Kalman Priors.

Dataset overview

Item	Value
Video sequences	1,579
Regular sequences	1,487
Dedicated occlusion sequences	92
Total frames	157,900
Frames per sequence	100
Resolution	1024 x 1024
Frame rate	20 fps
Duration per sequence	5 seconds
Video codec	H.264
Frames with a bounding box	151,006
Frames with a usable bounding box	145,902
Approximate download size	3.86 GiB

A usable bounding box is present and is not covered by an annotation-level invalid marker.

Conditions

Illumination

Lighting	Sequences	Frames
Day	942	94,200
Dusk	316	31,600
Night	321	32,100

Camera tilt

Camera tilt	Sequences	Frames
0 degrees	955	95,500
10 degrees	314	31,400
20 degrees	310	31,000

All dusk and night sequences use a camera tilt of 0 degrees.

Camera-trajectory region

Right, Left, Bottom, and Top identify four predefined camera-trajectory regions in MatrixCity. They do not describe the target's facing direction.

Region	Sequences	Frames
Bottom	354	35,400
Left	395	39,500
Right	372	37,200
Top	458	45,800

Occlusion subset

The dedicated occlusion subset contains 92 sequences and 9,200 frames. Bounding-box gaps form 109 contiguous events, with a mean event length of 23.65 frames.

Data generation

As described in the associated paper, MVOT was generated with MatrixCity rendering scripts in Unreal Engine. Camera positions were sampled at regular intervals along predefined trajectories at a fixed height. Scene pitch and illumination were varied under controlled settings, and target bounding boxes were obtained through the Unreal Engine actor-tracking API.

Directory structure

.
|-- README.md
|-- LICENSE
|-- CITATION.cff
|-- DATA_DICTIONARY.md
|-- annotations/
|   |-- regular/{lighting}/{camera_tilt}/{view}/*.csv
|   `-- occlusion/{lighting}/{camera_tilt}/{view}/*.csv
|-- data/
|   |-- metadata.jsonl
|   `-- videos/{subset}/{lighting}/{camera_tilt}/{view}/*.mp4
|-- manifests/
|   |-- checksums.sha256
|   |-- files.csv
|   `-- quality_report.json
`-- scripts/
    |-- upload_to_hub.py
    `-- validate_dataset.py

Loading the dataset

Install Hugging Face Datasets with video support, then load the repository:

from datasets import load_dataset

dataset = load_dataset(
    "Frank0666/MVOT",
    data_dir="data",
    split="train",
)

sample = dataset[0]
print(sample["sample_id"])
print(sample["annotations"][0])

For a local checkout:

from datasets import load_dataset

dataset = load_dataset("videofolder", data_dir="data", split="train")

The repository provides one default split. For model evaluation, define splits at the trajectory level where possible so that closely related clips do not appear in both training and evaluation sets.

Annotation format

Bounding boxes use pixel-space [x, y, width, height], where (x, y) is the top-left corner and the image origin is also at the top left. Each video record contains 100 ordered frame annotations. A missing box is represented by bbox_xywh: null and has_bbox: false.

Two frame-level fields support annotation filtering:

• has_bbox indicates that all four box coordinates are present.
• source_marked_invalid indicates that an invalid marker or range covers the frame.

For conservative training data selection, require has_bbox == true and source_marked_invalid == false.

The original frame-level occlusion labels are retained in occlusion_raw. Sequence membership in the dedicated occlusion subset is represented separately by subset == "occlusion".

See DATA_DICTIONARY.md for the complete metadata and annotation schema.

Validation

Run the structural validator from the repository root:

python scripts/validate_dataset.py
python scripts/validate_dataset.py --checksums

manifests/files.csv lists every video, annotation file, and metadata file in the release payload. manifests/checksums.sha256 provides their SHA-256 hashes.

Intended uses

MVOT supports single-object tracking, small-object localization, controlled robustness studies, and synthetic-to-real research across illumination, viewpoint, trajectory, and occlusion conditions.

Limitations

• Target category names are not encoded in the annotations.
• Synthetic scenes do not capture the full sensor, atmospheric, and viewing variability of real satellite imagery.
• The repository does not provide an official train, validation, and test assignment.

MVOT is intended as a controlled complement to real-world datasets, not as a replacement for them.

License

MVOT is released under the MIT License. See LICENSE for the full terms.

Citation

@article{fan2025satsam2,
  title   = {SatSAM2: Motion-Constrained Video Object Tracking in Satellite Imagery using Promptable SAM2 and Kalman Priors},
  author  = {Fan, Ruijie and Ye, Junyan and Chen, Huan and Huang, Zilong and Wang, Xiaolei and Li, Weijia},
  journal = {arXiv preprint arXiv:2511.18264},
  year    = {2025},
  doi     = {10.48550/arXiv.2511.18264}
}