n-iv commited on
Commit
bfbd233
·
verified ·
1 Parent(s): bb31f73

Camera-ready card: authors, radio definitions, unevaluated companion, NeurIPS citation

Browse files
Files changed (1) hide show
  1. README.md +21 -14
README.md CHANGED
@@ -35,6 +35,10 @@ configs:
35
 
36
  # V33DA++: Overlapping-Call and Long-Context Companion to V33DA
37
 
 
 
 
 
38
  V33DA++ is the companion release to the [V33DA](https://huggingface.co/datasets/songbirdini/V33DA) benchmark. It is built from the same BirdPark recordings and the same 10 zebra finches, but covers data **outside** the strict single-vocalizer attribution clips that define V33DA. V33DA++ is intended for tasks that benefit from **overlapping callers** or **longer temporal context**: source separation, vocal activity detection, audio-visual synchronization, and active-speaker detection.
39
 
40
  V33DA++ ships **three buckets**, each as a separate config:
@@ -45,7 +49,7 @@ V33DA++ ships **three buckets**, each as a separate config:
45
  | `padded` | 33,625 | ±2 s context windows around V33DA single-vocalizer events. Original event onsets and offsets are preserved inside the longer window, supplying long-range context without altering the underlying call. |
46
  | `overlap_padded` | 3,784 | The `overlap` bucket with the same ±2 s context, for long-context multi-caller tasks. |
47
 
48
- > V33DA++ is **not** the benchmark used in the V33DA paper's main results tables. For the single-vocalizer attribution benchmark, use [V33DA](https://huggingface.co/datasets/songbirdini/V33DA).
49
 
50
  **Code:** [github.com/marisbasha/v33da](https://github.com/marisbasha/v33da)  ·  **Main benchmark:** [songbirdini/V33DA](https://huggingface.co/datasets/songbirdini/V33DA)
51
 
@@ -148,7 +152,7 @@ Each parquet row is one vocalization event. Most binary array columns are NumPy
148
  |---|---|---|
149
  | `audio` | bytes | NumPy-serialized `(5, T)` float32 array — 5 cage microphones at 24,414 Hz, padded by `context_pad_sec` on each side |
150
  | `audio_path` | string | Path to a separately exported multichannel WAV mirror (best-effort, see Release contents) |
151
- | `accelerometer` | bytes | NumPy-serialized `(N_birds, T)` float32 array — per-bird on-body vibration channel at 24,414 Hz, used as the verification oracle |
152
  | `accelerometer_path` | string | Path to a separately exported multichannel WAV mirror (best-effort) |
153
  | `video_path` | string | Path to a composite multi-view MP4 (top, back, side) — `null` if not available |
154
 
@@ -166,16 +170,19 @@ Per-frame indexing: bird `i`'s keypoint `k` at frame `t` is `kp3d[t, i, k]` (and
166
 
167
  ### Radio telemetry (every visible bird, every frame)
168
 
169
- Per-bird FM radio signals from a 4-antenna receiver array (4 antennas: A, B, C, D; M = mean over antennas). Each is a NumPy `(N_birds, T_radio)` float32 byte blob.
170
 
171
  | Field | Description |
172
  |---|---|
173
- | `radio_frq` | Per-bird FM carrier frequency offset (Hz) |
174
- | `radio_thetaS` | Source-side phase used for triangulation |
175
- | `radio_thetaA / B / C / D` | Per-antenna received phase (rad) |
176
- | `radio_phiA / B / C / D / M` | Per-antenna and mean phase difference (rad) |
177
- | `radio_powRA / B / C / D / M` | Per-antenna and mean received power (linear) |
178
- | `radio_powNA / B / C / D / M` | Per-antenna and mean noise floor (linear) |
 
 
 
179
 
180
  ### Overlap-event metadata (`overlap` and `overlap_padded` only)
181
 
@@ -233,10 +240,10 @@ CC-BY-4.0. External tools used during preprocessing (e.g., Whisper / WhisperSeg,
233
  If you use V33DA or V33DA++, please cite the V33DA paper:
234
 
235
  ```bibtex
236
- @unpublished{basha2026v33da,
237
- title = {Who Called? V33DA: A Physically Verified Multimodal Benchmark for Vocal Attribution in Zebra Finch Groups},
238
- author = {Basha, Maris and Wang, Yuhang and Chen, Xiaoran and Cheng, Longbiao and Yapura, Luca and Zai, Anja T. and Salzmann, Mathieu and Hahnloser, Richard},
239
- year = {2026},
240
- note = {Under review},
241
  }
242
  ```
 
35
 
36
  # V33DA++: Overlapping-Call and Long-Context Companion to V33DA
37
 
38
+ **Authors:** Maris Basha¹, Yuhang Wang¹, Xiaoran Chen², Longbiao Cheng¹, Luca Yapura¹, Anja T. Zai¹, Mathieu Salzmann², Richard Hahnloser¹
39
+ ¹ Institute of Neuroinformatics, University of Zurich and ETH Zurich · ² Swiss Data Science Center and CVLab, EPFL
40
+ **Maintainer:** Maris Basha · **Provenance:** same BirdPark recordings, pipeline, and animal license (ZH148/2020) as V33DA.
41
+
42
  V33DA++ is the companion release to the [V33DA](https://huggingface.co/datasets/songbirdini/V33DA) benchmark. It is built from the same BirdPark recordings and the same 10 zebra finches, but covers data **outside** the strict single-vocalizer attribution clips that define V33DA. V33DA++ is intended for tasks that benefit from **overlapping callers** or **longer temporal context**: source separation, vocal activity detection, audio-visual synchronization, and active-speaker detection.
43
 
44
  V33DA++ ships **three buckets**, each as a separate config:
 
49
  | `padded` | 33,625 | ±2 s context windows around V33DA single-vocalizer events. Original event onsets and offsets are preserved inside the longer window, supplying long-range context without altering the underlying call. |
50
  | `overlap_padded` | 3,784 | The `overlap` bucket with the same ±2 s context, for long-context multi-caller tasks. |
51
 
52
+ > V33DA++ is an **unevaluated companion release**: no result in the V33DA paper depends on it. For the single-vocalizer attribution benchmark, use [V33DA](https://huggingface.co/datasets/songbirdini/V33DA).
53
 
54
  **Code:** [github.com/marisbasha/v33da](https://github.com/marisbasha/v33da)  ·  **Main benchmark:** [songbirdini/V33DA](https://huggingface.co/datasets/songbirdini/V33DA)
55
 
 
152
  |---|---|---|
153
  | `audio` | bytes | NumPy-serialized `(5, T)` float32 array — 5 cage microphones at 24,414 Hz, padded by `context_pad_sec` on each side |
154
  | `audio_path` | string | Path to a separately exported multichannel WAV mirror (best-effort, see Release contents) |
155
+ | `accelerometer` | bytes | NumPy-serialized `(N_birds, T)` float32 array — per-bird on-body vibration channel at 24,414 Hz, used to verify the caller; withheld from benchmark inputs |
156
  | `accelerometer_path` | string | Path to a separately exported multichannel WAV mirror (best-effort) |
157
  | `video_path` | string | Path to a composite multi-view MP4 (top, back, side) — `null` if not available |
158
 
 
170
 
171
  ### Radio telemetry (every visible bird, every frame)
172
 
173
+ Per-bird BirdRadio signals from a 4-antenna receiver (antennas A–D; M = master vector). Each is a NumPy `(N_birds, T_radio)` float32 byte blob with one value per radio frame (512 samples at 24,414 Hz = one video frame). Symbols follow the BirdPark paper (Rüttimann et al., PeerJ 2025); see the [V33DA card](https://huggingface.co/datasets/songbirdini/v33da#radio-telemetry).
174
 
175
  | Field | Description |
176
  |---|---|
177
+ | `radio_frq` | PLL tracking frequency |
178
+ | `radio_thetaS` | Master-vector phase angle $\theta_i$ (deg) |
179
+ | `radio_thetaA / B / C / D` | Per-antenna rotated-vector angles $\alpha_i^a$ (deg) |
180
+ | `radio_phiA / B / C / D` | Per-antenna compensation phases $\phi_i^a$ (deg) |
181
+ | `radio_phiM` | Sum of the compensation phases (deg) |
182
+ | `radio_powRA / B / C / D / M` | Baseband signal power (dBFS) |
183
+ | `radio_powNA / B / C / D / M` | Baseband signal variance (dBFS) |
184
+
185
+ BirdRadio does not compute an angle of arrival.
186
 
187
  ### Overlap-event metadata (`overlap` and `overlap_padded` only)
188
 
 
240
  If you use V33DA or V33DA++, please cite the V33DA paper:
241
 
242
  ```bibtex
243
+ @inproceedings{basha2026v33da,
244
+ title = {Who Called? V33DA: A Physically Verified Multimodal Benchmark for Vocal Attribution in Zebra Finch Groups},
245
+ author = {Basha, Maris and Wang, Yuhang and Chen, Xiaoran and Cheng, Longbiao and Yapura, Luca and Zai, Anja T. and Salzmann, Mathieu and Hahnloser, Richard},
246
+ booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
247
+ year = {2026},
248
  }
249
  ```