mcos-decoder

Decode MATLAB MCOS objects (Computer Vision Toolbox groundTruth, Audio/Signal Labeler exports, etc.) from .mat files in pure Python — no MATLAB license required.

Why this exists

MATLAB's Computer Vision Toolbox saves video annotations as a groundTruth MCOS object. When loaded with scipy.io.loadmat (or pymatreader, mat73), the actual labels are hidden inside an opaque __function_workspace__ blob:

>>> import scipy.io
>>> m = scipy.io.loadmat("IR_DRONE_001_LABELS.mat")
>>> m["None"]
array([(b'gTruth', b'MCOS', b'groundTruth', array([[3707764736], ...])), ...])
# ↑ no label data accessible

This problem has been open in scipy since 2024 and is the subject of multiple unanswered help threads (e.g. DroneDetectionThesis Issue #3). Until now the only third-party reader was the C# library MatFileHandler.

mcos-decoder fills the gap for Python users by walking the __function_workspace__ MAT5 element tree and extracting per-frame label cells.

Install

pip install mcos-decoder

Quick start

from mcos_decoder import load_groundtruth

bboxes = load_groundtruth("IR_DRONE_001_LABELS.mat")
# → list[tuple | None] of length n_frames
# Each entry is (x, y, w, h) for filled frames, or None when target absent.

for frame_idx, bbox in enumerate(bboxes, start=1):
    if bbox is None:
        print(f"Frame {frame_idx}: target absent")
    else:
        x, y, w, h = bbox
        print(f"Frame {frame_idx}: bbox=({x:.1f}, {y:.1f}, {w:.1f}, {h:.1f})")

Lower-level API

If your .mat uses a non-default MCOS layout (e.g. older MATLAB versions), you can override the depth and slot sizes:

from mcos_decoder import load_groundtruth

# Audio Labeler often nests one level deeper
bboxes = load_groundtruth(
    "audio_labels.mat",
    frame_depth=6,      # default 5
    bbox_size=80,       # default 80 (filled cell size in bytes)
    empty_size=112,     # default 112 (empty marker size)
)

For raw element-tree access:

import scipy.io
from mcos_decoder import MatStream

m = scipy.io.loadmat("labels.mat")
fw = m["__function_workspace__"].tobytes()
stream = MatStream(fw, skip_header=8)

for elem in stream.walk():
    if elem.type_name == "miMATRIX":
        print(f"depth-1 matrix at offset {elem.offset:#x}, "
              f"size {elem.payload_size}, "
              f"children: {len(elem.children)}")

Verified compatibility

Source	MATLAB version	Status
Halmstad Drone Detection Dataset (Svanström 2021)	R2020a	✓ 365 IR videos, 4 classes
Anti-UAV410 retrained tracker results	—	N/A (uses txt)

If you successfully use the package on another dataset, please open a PR adding it to this table.

Limitations

• Single-target per cell. The current decoder returns one bbox per frame; if your .mat has multiple bboxes per frame (e.g. MOT-style multi-object labels), the first one is returned. Multi-target support is on the roadmap.
• Numeric bbox only. Categorical labels are not yet extracted.
• groundTruth only. Other MCOS types (e.g. videoLabeler) may require different frame_depth / size parameters.

How it works

The MCOS object data is stored in __function_workspace__ as a recursive MAT5 element stream. Each element has an 8-byte tag (type, size); type 14 (miMATRIX) elements contain nested streams. Per-frame label cells are serialized at a fixed nesting depth (5 in current MATLAB versions) with size 80 (filled) or 112 (empty marker). The decoder walks the tree breadth-first and extracts the bbox doubles from the filled cells.

See mcos_decoder/groundtruth.py for the full implementation (~70 lines).

Citation

If this package supports your research, please cite:

@software{ozdemir2026mcos,
  author    = {Ozdemir, Burak},
  title     = {mcos-decoder: A Pure-Python Reader for MATLAB Computer
               Vision Toolbox groundTruth Objects},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19728531},
  url       = {https://github.com/bozdemir/mcos-decoder}
}

License

MIT — see LICENSE.