Update dataset card

README.md

@@ -14,6 +14,8 @@ tags:
 - attribution
 - radio-telemetry
 - accelerometer
+- zebra-finch
+- benchmark
 size_categories:
 - 10K<n<100K
 configs:
@@ -23,115 +25,124 @@ configs:
     path: v33da-*.parquet
 ---
 
-# V33DA: Benchmarking Vocal Attribution in Freely-Behaving Zebra Finches
+# V33DA: A Multi-Modal Benchmark for Vocal Attribution in Freely Interacting Zebra Finches
 
-V33DA is a multimodal benchmark for vocal attribution and 3D vocal localization in small groups of freely behaving zebra finches recorded in the BirdPark aviary. Each released sample is a single vocalization event with synchronized microphone audio, video, 3D pose, radio telemetry, and the accelerometer-derived on-body verification signal used to establish the ground-truth caller label.
+**Task.** Given a detected zebra finch vocalization and the set of birds visible at that moment, determine which bird produced the call.
 
-The dataset is designed so that caller labels do not reduce to the same room-acoustic attribution problem solved by the benchmark models. Instead, the release follows the paper’s extraction pipeline: candidate events are proposed from the body-mounted verification channel, manually reviewed, and then filtered by geometric and motion consistency before inclusion.
+V33DA provides 33,625 vocalization events from 10 individually identified zebra finches across 3 experiments (2021--2023), each with synchronized 5-channel audio, multi-view video, 3D pose, FM radio telemetry, and accelerometer-derived ground-truth labels.
 
-## Paper Summary
+| | BP01 | BP02 | BP05 | Total |
+|---|---|---|---|---|
+| Birds | 4 | 3 | 3 | 10 |
+| Events | 16,219 | 11,577 | 5,829 | 33,625 |
+| Recording days | 10 | 3 | 3 | 16 |
+| Raw hours | ~46h | ~14h | ~14h | ~74h |
 
-Recordings were collected from groups of 3--4 zebra finches during natural social interactions. The sensing setup aligns the same event across:
+**Code:** [github.com/marisbasha/v33da](https://github.com/marisbasha/v33da)
 
-- five cage microphones sampled at 24,414 Hz
-- three synchronized camera views
-- five 3D body keypoints per bird
-- 21 per-bird radio telemetry signals per frame
-- the raw accelerometer-derived vibration signal carried by each bird
+## Quick start
 
-The benchmark supports a task introduced in the paper:
-
-1. vocal attribution: identify which visible bird produced the vocalization
-
-## Modalities
-
-Each parquet row corresponds to one released vocalization event and includes:
-
-- `audio_path`: relative path to the multichannel cage-microphone WAV
-- `accelerometer_path`: relative path to the multichannel accelerometer WAV
-- `video_path`: relative path to the aligned composite MP4 clip
-- `keypoints_3d`: triangulated 3D body pose
-- `keypoints_2d_top`: 2D keypoints in the top-view image space
-- `keypoints_2d_back`: 2D keypoints in the back-view image space
-- `radio_*`: synchronized per-bird radio telemetry arrays
-- `bird_color`: color identity used as the source of truth for caller assignment
-- `vocalizer_idx`: index of the vocalizing bird
-
-Media is stored as external files and referenced from parquet by path.
-
-## Labeling and Filtering
-
-The paper’s released set is built from a reviewed candidate pool. After manual verification, samples are retained only if they satisfy the release filters:
-
-- overlapping vocalizations removed
-- 3D-to-2D reprojection error `<= 40 px`
-- maximum frame-to-frame 2D displacement `<= 40 px`
-- no NaNs in released 3D or 2D keypoints
-
-Color is the source of truth for the released vocalizer identity.
-
-## Release Contents
-
-This standalone release contains:
-
-- `v33da-*.parquet`: pooled parquet shards for the full release
-- `audio/`: multichannel microphone WAV files
-- `accelerometer/`: multichannel accelerometer WAV files
-- `clips/`: aligned composite MP4 clips
-- `calibrations/`: per-experiment camera calibration files
-- `explore.ipynb`: notebook for dataset exploration and reprojection
-- `metadata.json`: release schema and filter metadata
-
-## Dataset Size
-
-- released samples: `33,625`
-- pooled parquet shards: `23`
-- experiments: `3`
-- individuals: `11`
+```bash
+pip install huggingface_hub
+huggingface-cli download songbirdini/V33DA --repo-type dataset --local-dir data/v33da
+```
 
-Per-group counts in this standalone release:
+Or in Python:
 
-- `juvExpBP01 / blue`: `5,735`
-- `juvExpBP01 / peach`: `2,277`
-- `juvExpBP01 / red`: `3,963`
-- `juvExpBP01 / white`: `4,244`
-- `juvExpBP02 / blue`: `4,923`
-- `juvExpBP02 / peach`: `4,589`
-- `juvExpBP02 / red`: `2,065`
-- `juvExpBP05 / brown`: `3,633`
-- `juvExpBP05 / purple`: `2,140`
-- `juvExpBP05 / yellow`: `56`
+```python
+from huggingface_hub import snapshot_download
+snapshot_download("songbirdini/V33DA", repo_type="dataset", local_dir="data/v33da")
+```
 
-## Using the Dataset
+### Loading a sample
 
-The canonical release format is the pooled `v33da-*.parquet` set. Each binary array column is serialized with `numpy.save`, so decoding is:
+Each binary array column is serialized with `numpy.save`. To decode:
 
 ```python
 import io
 import numpy as np
+import pyarrow.parquet as pq
 
-arr = np.load(io.BytesIO(raw_bytes))
-```
+table = pq.read_table("data/v33da/v33da-00000.parquet")
+row = table.to_pydict()
 
-The included `explore.ipynb` notebook shows how to:
+# 3D keypoints: (N_birds, 5, 3) float64
+kp3d = np.load(io.BytesIO(row["keypoints_3d"][0]))
 
-- load parquet shards
-- inspect audio, accelerometer, video, pose, and radio
-- use the shipped camera calibrations to reproject 3D keypoints into 2D views
+# Audio path -> multichannel WAV
+audio_path = row["audio_path"][0]  # e.g. "audio/juvExpBP01/2021-06-28/..."
+```
 
-## Limitations
+The included `explore.ipynb` notebook shows how to inspect audio, video, pose, radio, and accelerometer data, and how to reproject 3D keypoints into 2D camera views using the shipped calibrations.
 
-As discussed in the paper, V33DA is collected in one aviary with one recording geometry across three experiments. Cross-experiment evaluation therefore combines bird-identity shift, group-composition shift, and recording-date shift. The released benchmark also excludes overlapping vocalizations and filters out samples with poor geometric consistency or excessive motion.
+## Modalities
 
-## Citation
+Each parquet row is one vocalization event:
+
+| Column | Description |
+|---|---|
+| `audio_path` | 5-channel cage-microphone WAV (24,414 Hz, float32) |
+| `accelerometer_path` | Per-bird accelerometer WAV (oracle verification signal) |
+| `video_path` | Aligned composite MP4 clip (3 views, 47.68 fps) |
+| `keypoints_3d` | Triangulated 3D pose: 5 keypoints per bird (beak, head, backpack, tail base, tail tip) |
+| `keypoints_2d_top` | 2D keypoints in top-view image space |
+| `keypoints_2d_back` | 2D keypoints in back-view image space |
+| `radio_*` | 21 per-bird FM radio telemetry signals per frame |
+| `vocalizer_idx` | Ground-truth caller index |
+| `bird_color` | Color identity of the vocalizer |
+| `experiment` | Experiment name (juvExpBP01 / juvExpBP02 / juvExpBP05) |
+| `date` | Recording date |
+
+## Ground-truth labels
+
+Labels come from **on-body accelerometer vibration**, not from microphone-based localization. A WhisperSeg model proposes candidate windows from the demodulated signal; events are retained only when audible in microphones, showing characteristic on-body vibration, with no overlapping call from another bird. All candidates are manually reviewed.
+
+### Filtering
+
+From 39,329 reviewed candidates:
+- 1,768 rejected as invalid during manual review
+- 3,784 removed by overlap filtering (+/-10 ms)
+- 152 removed by 3D-to-2D reprojection error (> 40 px)
+
+Yielding **33,625 released events**.
+
+## Per-bird counts
+
+| Experiment | Bird | Events |
+|---|---|---|
+| BP01 | blue | 5,735 |
+| BP01 | peach | 2,277 |
+| BP01 | red | 3,963 |
+| BP01 | white | 4,244 |
+| BP02 | blue | 4,923 |
+| BP02 | peach | 4,589 |
+| BP02 | red | 2,065 |
+| BP05 | brown | 3,633 |
+| BP05 | purple | 2,140 |
+| BP05 | yellow | 56 |
+
+## Release contents
+
+| Path | Description |
+|---|---|
+| `v33da-*.parquet` | 23 pooled parquet shards |
+| `audio/` | Multichannel microphone WAV files |
+| `accelerometer/` | Multichannel accelerometer WAV files |
+| `clips/` | Aligned composite MP4 clips |
+| `calibrations/` | Per-experiment camera calibration files |
+| `explore.ipynb` | Dataset exploration notebook |
+| `metadata.json` | Release schema and filter metadata |
+
+## Evaluation regimes
+
+The [benchmark code](https://github.com/marisbasha/v33da) supports three evaluation regimes:
+
+| Regime | What transfers? | Use case |
+|---|---|---|
+| **Session-disjoint** | Same birds, different days | In-domain ceiling |
+| **Held-out experiment** | New birds, new year, new cage | Cross-cohort transfer |
+| **Leave-one-bird-out** | One unseen bird per fold | Identity-shift stress test |
 
-If you use this dataset, please cite the accompanying paper:
+## Limitations
 
-```bibtex
-@inproceedings{basha2026v33da,
-  title={V33DA: Benchmarking Vocal Attribution in Freely-Behaving Zebra Finches},
-  author={Basha, Maris and Wang, Yuhang and Chen, Xiaoran and Cheng, Longbiao and Yapura, Luca and Salzmann, Mathieu and Hahnloser, Richard},
-  booktitle={XXX},
-  year={2026}
-}
-```
+V33DA is collected in one aviary with one recording geometry across three experiments. Cross-experiment evaluation combines bird-identity shift, group-composition shift, and recording-date shift. The released benchmark excludes overlapping vocalizations and filters out samples with poor geometric consistency.