MAAP3D Model Zoo
Pretrained detection and keypoint models for MAAP3D, hosted in MAAP3D/ModelZoo. The zoo includes six pose config/checkpoint pairs and four detector checkpoints. Supplied models are drafts; inference has not been validated for this release.
Model slots
| Model ID | Folder | Task | Status |
|---|---|---|---|
pigeon-detection |
Pigeon detection | Bounding-box detection | Draft; files supplied, inference unvalidated |
pigeon-pose |
Pigeon pose | Keypoint estimation | Draft; files supplied, inference unvalidated |
jay-detection |
Jay detection | Bounding-box detection | Draft; files supplied, inference unvalidated |
jay-pose |
Jay pose | Keypoint estimation | Draft; files supplied, inference unvalidated |
crow-detection |
Crow detection | Bounding-box detection | Draft; files supplied, inference unvalidated |
crow-pose |
Crow pose | Keypoint estimation | Draft; files supplied, inference unvalidated |
bird-detection |
Not supplied | Bounding-box detection | No detector supplied |
bird-pose |
Bird pose | Keypoint estimation | Draft; files supplied, inference unvalidated |
ape-detection |
Ape detection | Bounding-box detection | Draft; files supplied, inference unvalidated |
ape-pose |
Ape pose | Keypoint estimation | Draft; files supplied, inference unvalidated |
greti-detection |
Greti detection | Bounding-box detection | Template; detector not supplied |
greti-pose |
Greti pose | Keypoint estimation | Draft; files supplied, inference unvalidated |
Species coverage and training sources are supplied by the model author and documented in the individual model cards. Pose technical metadata is extracted from the supplied configs; detector architecture and class labels remain unconfirmed. Qualitative performance statements are author-reported.
Model groups
| Group | Author-provided description / coverage | Requested citations |
|---|---|---|
| ape | Generalized ape model. Coverage: chimpanzees, bonobos. | MAAP3D; ChimpACT and OpenApePose for pose |
| bird | Generalized bird model. Coverage: a range of birds. | MAAP3D; CUB (matching release) and Animal Kingdom for pose |
| crow | Model trained on crow data. Coverage: large-billed crows. | MAAP3D |
| greti | Great tit and blue tit model. Coverage: great tits, blue tits. | MAAP3D, 3D-SOCS |
| jays | Siberian jay model family trained on the CHIRP dataset. Coverage: Siberian jays. | MAAP3D, CHIRP |
| pigeons | Pigeon model family trained on the 3D-POP dataset. Coverage: pigeons. | MAAP3D, 3D-POP |
Training histories in these cards describe the author-provided pose training sources. Detector-specific history is recorded separately where supplied.
License and citation policy
The author-selected default is AGPL-3.0-or-later for original MAAP3D model contributions, with upstream model/code/config/dataset terms retained. The complete artifacts' upstream licensing review is pending. See LICENSE_POLICY.md, LICENSE, and THIRD_PARTY_NOTICES.md. OpenMMLab-derived configs retain Apache-2.0 terms. The root card uses a mixed-license policy label; it does not assert a uniform verified license for every artifact.
Please cite MAAP3D for every model. For ape pose, also cite ChimpACT and OpenApePose; for bird pose, also cite the matching CUB release and Animal Kingdom. Continue to cite 3D-SOCS for Greti, CHIRP for jays, and 3D-POP for pigeons. See CITATIONS.md and CITATIONS.bib. These are scholarly citation requests.
Filenames and preparation status
Artifact names use the model ID as their stem, retaining the native extension:
pigeon-detection.pt, pigeon-pose.pth, and pigeon-pose.py, for example.
Each model's metadata records its original filenames, byte counts and SHA-256
hashes. Config and checkpoint contents are preserved.
template: no model files supplied for this slot.draft: files supplied, but loading, predictions and detector/pose pairing have not been validated. Metadata may still be incomplete.ready: files, required metadata and inference validation are complete.
No bird or greti detector checkpoint is supplied. All six supplied pose configs
specify top-down estimators requiring detection boxes. For these two pose models,
detector_requirement: "required" with recommended_detector: null means no
compatible detector has been supplied yet; it does not mean detection is optional.
Add your models
- Drop each inference checkpoint into its matching task folder. Use the model
ID as the filename stem and retain its native extension, such as
pigeon-detection.ptorpigeon-pose.pth. Name a pose configpigeon-pose.py. - For MMPose, include the actual inference config and every local config it
references. Remove dependencies on training-machine paths. A YOLO checkpoint
that embeds its configuration can leave
files.configasnull. - Edit that folder's
metadata.json: set the weights/config paths and complete the species, architecture, framework, labels, model license and other known details. Paths are relative to the folder containing that metadata file. - Complete the model's README, including intended use, training data, evaluation, limitations and an inference example using the actual files.
- Verify loading and inference through MAAP3D in the documented environment.
Use
status: "draft"once files are supplied. After required metadata and inference validation are complete, change it to"ready"and update the table. This is a preparation convention, not an implemented downloader API.
Do not create empty .pt, .pth or Python config placeholders. Missing paths
are represented by JSON null; unknown values should stay unknown.
Metadata conventions
Each model folder contains one metadata.json with schema_version: 1.
This is an initial descriptive format; MAAP3D does not currently consume it.
model_id: stable, unique identifier from the table above. When adding a second architecture for the same task, give it a distinct ID and subfolder.files.weights/files.config: relative paths to the actual checkpoint and main config. Usenullwhile missing or when a separate config is unnecessary.files.additional: relative paths to any other files required for inference, including local base configs. Keep all model dependencies in that model folder.class_names: names in the checkpoint's numeric class-ID order; do not infer them from the folder name.keypoint_names: names in exact checkpoint output order.skeletonandflip_pairscontain pairs of those names, not numeric indices. Define anatomical left/right consistently and ensure the inference config agrees.recommended_detector: a model ID for a compatible top-down detector. The provisional pose slots point to the same-group detector slot; compatibility has not been tested. For top-down models,detector_requirement: "required"distinguishes a missing detector from a model that needs no separate detector.inference.input_size:[width, height]in pixels when known. Thresholds are model-specific; no recommended threshold is assumed.compatibility: versions actually tested, not an unverified support promise. Record package names and version strings underpackages.evaluation.metrics: named metric/value entries. Document units, thresholds, evaluation split and procedure in the README; leave the object empty until measured. Split held-out data by individual/session as appropriate to the use case.license: terms for this model's weights. Do not assume the MAAP3D package license applies. Use an SPDX identifier when applicable, with a link or local license file giving the actual terms.
Remaining release metadata
- Exact base checkpoint identities, upstream grants and any separate permissions; specifically review ChimpACT pretraining for ape pose and the CUB version used for bird pose. Dataset terms do not automatically settle trained-weight terms.
- Detector architecture, ordered class labels and useful inference image sizes and confidence thresholds.
- Training sample/individual/session counts, exact data versions, training recipe and release/version label. Provide an exhaustive species list for the general bird model if available; scientific names can be added later.
- Held-out quantitative metrics, evaluation split definitions, known failure cases and tested environment versions, plus a small inference validation example.
- A preferred MAAP3D paper citation/DOI if available. A provisional software reference is already included.
These follow-ups do not erase the supplied author metadata or change the draft status. Checkpoint/config bytes and existing provenance hashes are preserved.
Downloads
Download checkpoints, configs and their metadata from the repository's Files and versions tab. Use a specific repository commit when downloading for reproducible analyses. Keep the selected model's documentation and the root license, policy, notices and citation files with your download.
A MAAP3D model-zoo download helper has not been implemented. Existing MAAP3D
inference functions can use downloaded local paths. MMPose models require both
the pose config and checkpoint. download_sample_data() continues to target
the separate SampleData repository.