SABER Dataset

This dataset accompanies the SABER paper:SABER enables identity-resolved quantification of multi-animal social behavior (in preparation).It contains the raw behavioral videos, the keypoint (pose estimation) datasets, and the frame-level behavior annotations used to develop and validate SABER, a framework for identity-resolved quantification of multi-animal social behavior from a single overhead video stream. Four social paradigms are covered:1.

Two-mouse spontaneous social behavior (40 videos, ~10 min each) — with frame-level behavior annotations for both animals (16 behavior classes).2. Four-mouse spontaneous social behavior (6 videos, ~10 min each).3. Aggressive behavior (17 videos, short fight clips, 22–76 s).4.

Mating behavior (7 videos, 23–113 s).In total the dataset contains 70 videos, 27,261 images with 27,261 keypoint label files, and 80 behavior annotation files.ContentsSABER_dataset/├── README.md├── video datasets/│  ├── 2-mouse social behavior/  40 AVI videos (2 mice, ~10 min each)│  ├── 4-mouse social behavior/   6 AVI videos (4 mice, ~10 min each)│  ├── Aggressive behavior/    17 MP4 videos (fight clips)│  └── Mating behavior/       7 MP4 videos├── behavior datasets/│  └── labels/           80 TXT files — frame-level behavior│                  annotations (Caltech Behavior│                  Annotator format), one per animal│                  per two-mouse video└── keypoints datasets/  ├── model_used_dataset/  │  └── dataset_060403_3000/  Pose-model training data  │    ├── mouse/       Body keypoints — 13,000 frames  │    │  ├── images/train/  12,191 JPG  │    │  ├── labels/train/  12,191 TXT  │    │  ├── images/val/    809 JPG  │    │  └── labels/val/    809 TXT  │    └── tail/        Tail keypoints — 12,961 frames  │      ├── images/train/  12,152 JPG  │      ├── labels/train/  12,152 TXT  │      ├── images/val/    809 JPG  │      └── labels/val/    809 TXT  └── validation_only_datasets/  Held-out pose validation sets    ├── 2-mouse social behavior_val/ 200 frames (images/ + labels/)    ├── 4-mouse social behavior_val/ 491 frames (images/val/ + labels/val/)    ├── Aggressive behavior_val/   445 frames (images/val/ + labels/val/)    └── Mating behavior_val/     164 frames (images/val/ + labels/val/)1.

Video datasetsvideo datasets/ contains the original behavioral recordings, grouped by paradigm.Folder             Format Videos Recording details2-mouse social behavior/    AVI   40   Two freely interacting mice in a 42 × 42 × 30 cm open-field arena, ~10 min per video. Files are named X-Y.avi (group X = 1–10, video Y = 1–4).4-mouse social behavior/    AVI   6    Four mice in a 27 × 39 × 27.5 cm arena, ~10 min per video.

Files are named by recording date (e.g., 20260110_4.avi).Aggressive behavior/      MP4   17   Aggressive encounters (fight clips, 22–76 s), e.g., 0.mp4–5.mp4, fight_1-1.mp4–fight_2-6.mp4.Mating behavior/        MP4   7    Male–female mating sessions (23–113 s), e.g., 6.mp4, sex_1-1.mp4.All videos were recorded from above at up to 1920 × 1080 px and 30 frames per second (see Methods below).

Frame indices in the annotation files can be converted to time by dividing by 30.2. Behavior datasets (frame-level behavior annotations)behavior datasets/labels/ contains 80 annotation files covering every animal of the 40 two-mouse social behavior videos: file X-Y_Z.txt annotates mouse Z (1 or 2) of video X-Y.avi in video datasets/2-mouse social behavior/.FormatFiles follow the Caltech Behavior Annotator text format. Each file has a short header declaring the behavior-key configuration, followed by an S1: table with one row per behavior bout:Caltech Behavior Annotator - Annotation FileS1: start end type-----------------------------  1 151 explore  152 193 stand...- start / end: first and last frame index (1-based) of the bout (30 fps).- type: behavior label (see table below).- Consecutive unlabeled frames belong to blank.3.

Keypoints datasets3.1 Pose-model training data — keypoints datasets/model_used_dataset/dataset_060403_3000/Training/validation data for the two SABER pose models (TMP):- mouse/ — body keypoints: 6 keypoints per animal instance.- tail/ — tail keypoints: 3 keypoints per animal instance.Each folder contains paired images/ (JPG frames, 1280 × 720) and labels/ (TXT) directories split into train/ and val/.

Every image has a same-named label file. File names encode the source video and frame index (e.g., 0_1015.jpg = video 0, frame 1015).3.2 Held-out pose validation sets — keypoints datasets/validation_only_datasets/Four paradigm-specific validation sets that were not used for model training, in the same YOLO keypoint format (6-keypoint body labels, one line per animal instance, all animals present in the frame):Folder              Frames Layout2-mouse social behavior_val/   200   images/ + labels/4-mouse social behavior_val/   491   images/val/ + labels/val/Aggressive behavior_val/     445   images/val/ + labels/val/Mating behavior_val/       164   images/val/ + labels/val/Note: the three *.cache files inside these folders are YOLO dataloader cache artifacts and can be ignored or deleted.3.3 Ultralytics keypoint label formatAll keypoint labels follow the Ultralytics keypoint format.

Each line describes one animal instance:<class> <cx> <cy> <w> <h> <kx1> <ky1> <v1> <kx2> <ky2> <v2> ... <kxn> <kyn> <vn>- class: always 0 (mouse for the body model, tail instance for the tail model).- cx, cy, w, h: bounding box center, width and height, normalized to [0, 1] relative to image size.- kx, ky: keypoint coordinates, normalized to [0, 1].- v: visibility flag — 2 visible, 1 occluded, 0 unlabeled (the three fields are then 0 0 0).Body keypoints (mouse/ and the validation sets), 6 per instance in this order:0 snout   1 head_center  2 body_center3 tailbase  4 ear_left    5 ear_rightTail keypoints (tail/), 3 per instance:0 tail_mid1  1 tail_mid2  2 tail_tip(the tail base is provided by the body model's keypoint 3).4.

Methods

Animals and ethicsMale and female C57BL/6J mice (8–12 weeks old) were purchased from the Laboratory Animal Center of Southern Medical University (license no. SCXK(Yue)2021-0041). All procedures followed the Guidelines for Ethical Review of Laboratory Animal Welfare (GB/T 35892-2018) and were approved by the Southern Medical University Animal Ethics Committee. Mice were housed under a 12 h light/dark cycle (lights on 08:00, off 20:00) at 22 ± 2 °C and 40–60 % humidity, ≤ 5 animals per cage, with food and water ad libitum.

Animals were handled twice daily for 3–4 days before experiments.Behavioral assays and video acquisition- Two-animal recordings (two-mouse social behavior, aggression, mating): mice moved freely in a white plastic open-field arena (42 × 42 × 30 cm).- Four-mouse recordings: a 27 × 39 × 27.5 cm arena.- Videos were acquired from above under white LED illumination (100–300 lux) with a KM2010-1080P camera (Kaimeng Vision) controlled by VideoCap software, at up to 1920 × 1080 px and 30 fps.- Animals acclimated to the arena for 30 min before recording.

Two- and four-mouse social sessions lasted ~10 min; aggression clips lasted 22–76 s; mating sessions 23–113 s.Behavior annotationFrame-level behavior labels were produced by trained annotators with the Caltech Behavior Annotator for every animal of the 40 two-mouse social behavior videos, using the 16-class ethogram listed in Section 2.Keypoint annotationKeypoint coordinates, bounding boxes, and identity/group information were annotated in LabelMe and exported as JSON.

A total of 3,000 randomly sampled frames were annotated — 1,000 from the two-mouse social behavior recordings, 1,400 from the aggressive behavior recordings, and 600 from the mating behavior recordings — with a 3:1 training-to-validation ratio. Annotations were subsequently converted to the Ultralytics keypoint format released here.5. Code and software- SABER (the associated analysis framework): github.com/kidous2333/SABER — released under an academic non-commercial license.- Project website: sunlabsaber.netlify.app- Videos: standard AVI/MP4 codecs, playable in any common media player.- Behavior annotations: plain-text tables (Caltech Behavior Annotator format), readable in any text editor or spreadsheet software, and compatible with the Caltech Behavior Annotator and similar behavioral annotation tools (e.g., BORIS).- Keypoint labels: Ultralytics keypoint format.

The dataset can be loaded directly with the ultralytics package; an example dataset configuration is provided in the SABER repository at config/datasets/TMP_mouse_dataset.yaml.6. Sharing / Access information- License: This dataset is released under the Creative Commons Zero (CC0 1.0) public domain dedication. You may copy, modify, and distribute the data, even for commercial purposes, without asking permission.-

Citation

If you use this dataset, please cite the accompanying paper: SABER enables identity-resolved quantification of multi-animal social behavior (in preparation).- Funding: This work was supported in part by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2023ZD0507100), the National Natural Science Foundation of China (grant no. 82271555, 82322024, and 82671967), and the Guangdong Major Project of Basic Research (2026B0303000010).- Contact: Xiang-Dong Sun (xisun@smu.edu.cn)