MC-Marmo

A computer-vision dataset of three common marmosets (Callithrix jacchus) living together in a single cage, recorded by two cameras that face each other. Every marmoset in every annotated frame carries three labels at once:

• an instance segmentation polygon that outlines the visible body.
• a tracking identity that is consistent across frames within one video.
• a posture label, one of five classes: climb, lay, sit, stand, or walk/jump.

Dataset detail

• 3 marmosets, unmarked and look alike, the same three in every video.
• 2 cameras (EZVIZ EB5 4k) at opposite ends of the cage, 180 degrees apart, 1280 x 720, recorded at about 15 fps and annotated at about 2 fps.
• 7 ten-minute sessions (set0-set6), each recorded by both cameras at the same time, over three days of daylight.
• 8,485 annotated frame pairs, 44,979 polygon masks, 37,910 marmoset appearances (a marmoset split into pieces by a perch counts once).
• 5 posture classes and their share of all masks: lay 46.2, climb 19.5, stand 14.6, walk/jump 10.5, sit 9.2 (percent).

Identities are assigned within one video only and are not matched across videos, so the same identity number in two different videos may be a different marmoset.

Train and test split

The dataset is divided into a train and test split that holds out a whole video. Training uses videos set1 to set6, and testing uses set0.

Folder layout

Raw videos and their annotation JSON, in a ready-to-use train and test split.

train/  annotations/set{1..6}_pair_annotations.json   videos/set{1..6}_cam{1,2}.mp4
test/   annotations/set0_pair_annotations.json         videos/set0_cam{1,2}.mp4

Loading example

import json
import cv2
from pathlib import Path

# images live inside the mp4 files.
root = Path("test")
doc = json.load(open(root / "annotations" / "set0_pair_annotations.json"))

cam1 = doc["annotations_a"]                                 # camera 1
cap = cv2.VideoCapture(str(root / "videos" / f"{cam1['video_id']}.mp4"))

for idx, frame in list(cam1["frames"].items())[:5]:
    cap.set(cv2.CAP_PROP_POS_FRAMES, int(idx))             # seek by frame index, not time
    ok, image = cap.read()                                 # image = frame `idx` of the mp4 (BGR)
    for poly in frame["polygons"]:
        print(cam1["video_id"], idx, poly["track_id"], poly["posture"],
              "points:", len(poly["points"]))
cap.release()

Annotation tool

The annotation web application is available at https://github.com/Makkawan/MC-Marmo-annotation-web-application.

Acknowledgment

The author gratefully acknowledges the Faculty of Medicine, Chulalongkorn University, for hosting the dataset, and the Anubis Marmoset Farm, Suphan Buri, for the daily care of the animals. This work was supported by the Department of Biomedical Engineering, Faculty of Engineering, Mahidol University.

License

Released under CC-BY-4.0. When using this dataset, please cite the paper below.

Citation

TBC