imageomics/mmla-pose

Dataset Card for MMLA Pose: Pose Orientation Labels for Zebras in MMLA Drone Footage

Cropped zebra images from MMLA drone footage, manually labeled with one of eight discrete pose orientations relative to the camera. The dataset was curated to train a pose / viewpoint classifier (DINOv2 backbone + MLP head) for the WildWing autonomous drone-navigation system, where knowing which side of an animal is visible drives both individual re-identification (matching against flank stripe patterns) and behavior-aware flight decisions (e.g., avoiding approaches from the front).

Figure 1. Pose classes reference diagram. The label indicates which flank of the zebra is visible to the camera, equivalently the direction the zebra's head is pointing relative to the camera. The dark circle in each icon marks the head. The camera/drone is positioned at the bottom of the diagram.

Dataset Details

This dataset contains 988 manually labeled 224x224 RGB crops of zebras observed in aerial drone footage from the Mpala Research Centre and Ol Pejeta Conservancy in Kenya, and at the Wilds Conservation Center in Ohio. Each crop is assigned exactly one of eight pose-orientation labels describing which flank of the zebra is visible to the camera (equivalently, the direction of the zebra's head relative to the camera). An additional 35 ambiguous or unusable crops are preserved in a _skip/ folder to document labeling decisions.

The labels were created to train the pose component of the WildWing drone-navigation stack. WildWing combines a YOLO-based animal detector with a pose classifier so the drone can reason about how each animal in its view is oriented and make better autonomous decisions about positioning. Reliable pose estimates are particularly important for individual re-identification of zebras, which depends on matching the lateral stripe pattern - so the classifier needs to know when a useful side view (left or right profile) is available versus when the animal is facing toward or away from the camera.

Supported Tasks and Leaderboards

Image classification (8-class): Predict the orientation of a zebra in a cropped aerial image. Labels are: front, front-left, left, back-left, back, back-right, right, front-right.

Suggested evaluation: top-1 accuracy with per-class breakdown; expect most confusion between adjacent classes (e.g., front vs. front-left). A practical aggregation is to group predictions into front-facing / side-profile / back-facing buckets, since downstream uses (e.g., re-ID) often care more about whether a usable flank is visible than about the exact 45-degree bin.

Dataset Structure

mmla-pose/
    _reference.png            # Visual diagram of the 8 pose classes
    front/                    # 49 images
    front-left/               # 84 images
    front-right/              # 44 images
    left/                     # 295 images
    right/                    # 369 images
    back-left/                # 56 images
    back-right/               # 44 images
    back/                     # 47 images
    _skip/                    # 35 images that were ambiguous or unusable;
                              #   excluded from training, kept for transparency

Data Instances

All images are 224x224 RGB JPEGs, cropped from frames of aerial drone footage. Each crop is placed directly into the folder for its pose label - the folder name is the label. Filenames encode provenance back to the source video and frame, with the format:

mpala_session_<S>_DJI_<VID>_partition_<P>_DJI_<VID>_<FRAME>_c<CROP_IDX>_<BBOX_ID>.jpg

For example, mpala_session_1_DJI_0002_partition_1_DJI_0002_000321_c1_000.jpg is from Mpala session 1, source video DJI_0002, partition 1, frame 321, crop channel 1, bounding box index 0.

Data Fields

The dataset is folder-based; there is no separate metadata file. The folder name is the label, and the filename uniquely identifies the source video and frame:

• Pose label: parent folder name, one of the 8 pose classes (or _skip for excluded crops).
• mpala_session_<S>: Mpala field-session identifier.
• DJI_<VID>: source DJI drone video file.
• partition_<P>: partition of the source video used during KABR preprocessing.
• <FRAME>: zero-padded frame number within the partition.
• c<CROP_IDX>: crop index for the frame.
• <BBOX_ID>: index of the bounding box (when multiple animals appear in the same frame).

Data Splits

No fixed train/val/test split is provided - the training script for the downstream classifier creates a random split at runtime. With only ~988 labeled images and a strongly imbalanced distribution (see below), users should expect to need stratified sampling and/or weighted sampling when training.

Class Distribution

Pose	# Images	Share
right	369	37.3%
left	295	29.9%
front-left	84	8.5%
back-left	56	5.7%
front	49	5.0%
back	47	4.8%
back-right	44	4.5%
front-right	44	4.5%
Total labeled	988	100%
_skip	35	(excluded)

The distribution is heavily skewed toward pure side profiles (left + right together account for ~67% of labels). This reflects the natural geometry of drone-side observation - side profiles are both the most common viewing angle and the most informative for stripe-based individual re-ID, so labelers prioritized them.

Dataset Creation

Curation Rationale

The crops were labeled specifically to train a pose / viewpoint classifier inside the WildWing drone-navigation system. Two downstream needs drove the schema:

1. Re-identification of individual zebras depends on matching lateral stripe patterns, so the system needs to know when a clean left- or right-side view is available.
2. Behavior-aware navigation - e.g., not approaching a zebra from directly in front - requires distinguishing front-facing, side, and rear orientations.

The 8-class scheme is therefore designed to be both coarse enough to label reliably from low-resolution drone crops and fine enough to support these two downstream uses.

Source Data

Data Collection and Processing

Source frames were drawn from the same drone footage used in the MMLA Wilds, MMLA Mpala, and MMLA Ol Pejeta Conservancy datasets. Animals were detected in each frame with a YOLO v11 detector trained for aerial wildlife imagery (see imageomics/mmla), each bounding box was cropped from the source frame, and the resulting crops were resized to 224x224 for compatibility with the DINOv2 backbone used in the downstream classifier. Crops were sampled across multiple Mpala sessions and source videos to cover varying lighting conditions, terrain, herd densities, and altitudes.

Who are the source data producers?

The underlying drone footage was collected by the KABR team at the Mpala Research Centre. See the KABR dataset card and paper for full provenance.

[Add OPC and Wilds]

Annotations

Annotation process

Each detected zebra crop was reviewed by a human labeler and placed into one of the 8 pose folders based on which flank is most visible to the camera, equivalently the direction the zebra's head is pointing relative to the camera. The _reference.png diagram (Figure 1 above) was used as the labeling guide. Crops that were too ambiguous to label confidently - typically heavy occlusion, motion blur, or only a small fraction of the animal visible - were placed in _skip/ and are excluded from the 8 pose-class totals.

Who are the annotators?

Labeling was performed by Claire as part of the WildWing project at the Imageomics Institute.

Personal and Sensitive Information

The dataset contains aerial imagery of wildlife only; no humans or identifiable personal information are included. Plains and Grevy's zebras are present in the imagery. Grevy's zebras (Equus grevyi) are classified as Endangered by the IUCN; however, the location information here is no more specific than "Mpala Research Centre", which is already public for the parent KABR dataset.

Considerations for Using the Data

• Small size. With ~988 labeled images, this dataset is intended for fine-tuning a frozen-backbone classifier (the WildWing pipeline freezes DINOv2 and trains only an MLP head). It is unlikely to be sufficient on its own for training a classifier from scratch.
• Class imbalance. left and right profiles dominate (~67% combined). Use stratified or weighted sampling, and report per-class metrics rather than overall accuracy alone.
• Horizontal-flip symmetry. Augmentation that flips an image horizontally must also flip the label: left <-> right, front-left <-> front-right, back-left <-> back-right; front and back are invariant. The WildWing training script handles this swap automatically.
• Adjacent-class confusion is expected. Pose orientation is continuous and the 8 bins are 45 degrees wide, so confusion between neighbors (e.g., front vs. front-left) is structural and not necessarily indicative of a model bug.
• _skip/ is not a class. The _skip/ folder documents what was excluded from labeling and should not be used as a 9th training class.

Bias, Risks, and Limitations

• Species coverage: All labeled crops are zebras (Plains and Grevy's, mixed). A classifier trained only on this data may not generalize to other species (e.g., giraffes) without additional labels.
• Geographic and seasonal coverage: All footage is from Mpala Research Centre, January 2023. Lighting and habitat conditions outside this scope are unrepresented.
• Aerial viewpoint: All crops are top-down/oblique aerial views; ground-level imagery is out of scope.
• Annotator coverage: Annotations were produced by a small labeling team; no inter-annotator agreement metric is computed for this release.

Recommendations

• Pair this dataset with the WildWing training script (see Imageomics/wildwing), which handles class balancing, the horizontal-flip label swap, and stratified splits.
• For downstream uses that only need "side view available vs. not", consider mapping the 8 labels into a smaller {front-facing, side, back-facing} schema before training.
• For deployment on additional species, collect and label species-specific crops rather than assuming the zebra-trained classifier transfers.

Licensing Information

This dataset is dedicated to the public domain under CC0 1.0 Universal. We ask that you cite the dataset (and the underlying KABR dataset) if you use it in your research.

Citation

BibTeX:

Data

@misc{mmla-pose-2026,
  author       = {Sun, Claire and Kline, Jenna},
  title        = {MMLA Pose: Pose Orientation Labels for Zebras in MMLA Drone Footage},
  year         = {2026},
  url          = {https://huggingface.co/datasets/imageomics/mmla-pose},
  publisher    = {Hugging Face},
  doi          = {<add once generated>}
}

Please also cite the underlying MMLA datasets (MMLA Wilds, MMLA Mpala, and MMLA Ol Pejeta Conservancy) and paper:

Underlying drone footage (MMLA)

@misc{kline2025mmla,
      title={MMLA: Multi-Environment, Multi-Species, Low-Altitude Drone Dataset}, 
      author={Jenna Kline and Samuel Stevens and Guy Maalouf and Camille Rondeau Saint-Jean and Dat Nguyen Ngoc and Majid Mirmehdi and David Guerin and Tilo Burghardt and Elzbieta Pastucha and Blair Costelloe and Matthew Watson and Thomas Richardson and Ulrik Pagh Schultz Lundquist},
      year={2025},
      eprint={2504.07744},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.07744}, 
}

Acknowledgements

This work was supported by the Imageomics Institute, which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under Award #2118240 (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Drone footage was collected at the Mpala Research Centre in Kenya as part of the broader KABR data collection effort. It was also collected at Ol Pejeta Conservancy in Kenya and at the Wilds Conservation Center in Ohio.

Glossary

• Pose / viewpoint: Which flank of the animal is visible to the camera, equivalently the direction the animal's head is pointing relative to the camera. Used interchangeably in this card.
• Crop: A 224x224 RGB sub-image extracted from a drone-video frame using a YOLO bounding box around a single animal.
• _skip: A folder of crops that were too ambiguous or occluded to label confidently. Not a training class.

More Information

For details on the downstream classifier architecture (DINOv2 + MLP head), training recipe, and how the pose predictions are consumed inside the autonomous drone-navigation stack, see the WildWing project and its repository.

Dataset Card Authors

Claire Sun and Jenna Kline (Imageomics Institute / WildWing project).

Dataset Card Contact

Open a discussion in the Community tab of this dataset.