README.md

---
pretty_name: SonoScene360
license: cc-by-4.0
language:
  - en
size_categories:
  - n<1K
annotations_creators:
  - expert-generated
source_datasets:
  - original
tags:
  - image
  - 3d
  - multimodal
  - spatial-audio
  - first-order-ambisonics
  - real-world
viewer: false
---

# SonoScene360: The SonoWorld Evaluation Dataset

This repository hosts **SonoScene360**, the real-world evaluation dataset introduced in **SonoWorld: From One Image to a 3D Audio-Visual Scene (CVPR 2026)**. SonoWorld generates a 3D audio-visual scene with spatialized sound from a single input image.

[[Paper](https://arxiv.org/abs/2603.28757)] [[Project website](https://humathe.github.io/sonoworld/)] [[Official code](https://github.com/HuMathe/sonoworld)]

The current dataset snapshot contains **10 scenes**, **34 calibrated microphone poses**, and **68 selected SN3D-normalized first-order Ambisonics (FOA) recordings**.

## Dataset statistics

| Item | Count or format |
| --- | --- |
| Real-world scenes | 10 |
| Calibrated microphone poses | 34 |
| Selected audio recordings | 68 |
| Original panoramas | 10 |
| Outpainted input panoramas | 10 |
| Rendered microphone-view panoramas | 34 |
| Audio format | SN3D FOA, 4 channels, 48 kHz, 16-bit PCM, approximately 10 seconds per recording |
| FOA channel order | W, Y, Z, X |

## Download

Download the complete repository with `huggingface_hub`:

```python
from huggingface_hub import snapshot_download

dataset_root = snapshot_download(
    repo_id="DerongJin/SonoScene360",
    repo_type="dataset",
)
```

Alternatively, use the Hugging Face CLI:

```bash
hf download DerongJin/SonoScene360 \
  --repo-type dataset \
  --local-dir SonoScene360
```

After downloading, use `data/metadata.json` as the dataset-wide index. All paths stored in the index are relative to the repository root.

## 1. Dataset contents

### Directory structure

```text
data/
|-- metadata.json
`-- <scene_id>/
    |-- audio/
    |   `-- <mic_id>/
    |       `-- <sample_id>/
    |           `-- foa.wav
    |-- images/
    |   |-- panorama.jpg
    |   |-- panorama_outpainted.jpg
    |   `-- <mic_id>/
    |       `-- rendered.jpg
    |-- metadata/
    |   |-- known_sources.json
    |   `-- <mic_id>/
    |       `-- metadata.json
    `-- video/                  # reserved; not included in the public release yet
```

Each scene provides the following data:

- `images/panorama_outpainted.jpg` is the outpainted equirectangular panorama used as the scene input to our method. `images/panorama.jpg` retains the original, non-outpainted panorama for reference.
- `audio/<mic_id>/<sample_id>/foa.wav` stores one audio sample. A microphone may have one or more selected samples. Every released file is approximately 10 seconds of four-channel, 48 kHz, 16-bit PCM, **SN3D-normalized FOA** audio in **W, Y, Z, X** channel order.
- `metadata/known_sources.json` stores the important semantic sound classes shared by the entire scene.
- `metadata/<mic_id>/metadata.json` stores the two 2D microphone annotations, microphone rotation, and text-labelled sound-source directions for one microphone pose. These fields are explained in [Microphone metadata](#microphone-metadata) and [Microphone calibration](#3-microphone-calibration).
- `images/<mic_id>/rendered.jpg` is a novel-view panorama rendered by our method at the corresponding microphone position. We use this image as the visual input for image-conditioned baselines because those baselines do not themselves render a novel view at the microphone position.

### Dataset-level index

`data/metadata.json` is the entry point for loading the release. All stored paths are relative to the repository root.

| Key | Type | Contents |
| --- | --- | --- |
| `audio_paths` | `list[str]` | All selected FOA paths, ordered by scene, microphone, and sample. |
| `scene_input_images_paths` | `dict[str, str]` | Scene-to-outpainted-model-input panorama mapping. |
| `scene_raw_images_path` | `dict[str, str]` | Scene-to-original, non-outpainted `panorama.jpg` mapping. |
| `scene_known_sources_paths` | `dict[str, str]` | Scene-to-`known_sources.json` mapping. |
| `scene_mic_metadata_paths` | `dict[str, dict[str, str]]` | Per-scene, per-microphone metadata paths. |
| `scene_mic_recording_paths` | `dict[str, dict[str, list[str]]]` | Per-scene, per-microphone lists of selected recordings. |
| `scene_names` | `list[str]` | All scene IDs. |
| `scene_sound_source_text_labels` | `dict[str, list[str]]` | Unique sound-source text labels associated with each scene. |

### Microphone metadata

Each `data/<scene_id>/metadata/<mic_id>/metadata.json` has the following form:

```json
{
  "2d_points": {
    "mic_center_2d": {"x": 8053, "y": 2929},
    "mic_location_2d": {"x": 8070, "y": 3571}
  },
  "apriltag": {
    "rotation": [
      [-0.626414, -0.777761, -0.051903],
      [0.027733, 0.044306, -0.998633],
      [0.778997, -0.626997, -0.006184]
    ]
  },
  "mic_id": "mic_08",
  "scene_id": "fountain-multi",
  "sound_source_annotations": [
    {"direction": "right", "text_label": "fountain"}
  ]
}
```

- `mic_location_2d` is the manually annotated ground sticker directly below the microphone.
- `mic_center_2d` is the manually annotated microphone center. It was called the elevation point during data collection because, together with the ground point, it determines the microphone elevation angle.
- Both points are integer pixel coordinates in the original calibration panorama: `x` increases to the right and `y` increases downward. They are not normalized coordinates.
- `apriltag.rotation` is a `3 x 3` world-to-microphone rotation matrix. Its exact convention is described below.
- `sound_source_annotations` contains a semantic label and one of `front`, `left`, `right`, or `behind` for each annotated source.

Each `data/<scene_id>/metadata/known_sources.json` is shared by every microphone in that scene:

```json
{
  "known_sources": ["fountain"]
}
```

The canonical scene-level values are:

| Scene | Known sources |
| --- | --- |
| `fountain-multi` | `fountain` |
| `kitchen-multi` | `faucet`, `microwave` |
| `pool` | `pool` |
| `pool-2` | `pool` |
| `river-bridge-river` | `stream` |
| `river-bridge-train` | `stream`, `train` |
| `stream` | `stream` |
| `stream-walk` | `stream`, `person walking on leaves` |
| `two-birds` | `birds chirping in leaves` |
| `two-birds-siren` | `police siren` |

### Example visualization

The following audit sheet shows `fountain-multi / mic_08 / sample_001`:

![Fountain-multi mic 08 data visualization](example/fountain-multi_mic-08_data-visualization.jpg)

The left panel uses `images/mic_08/rendered.jpg` as its background. Cyan shows the annotated, text-labelled source direction, while magenta shows the dominant direction estimated from the FOA signal. The right panel shows relative FOA directional energy, normalized per recording to 0 dB and clipped at -30 dB; its magenta point marks the same estimated peak.

For all equirectangular visualizations, image azimuth runs from **-180 degrees on the left to +180 degrees on the right**: left is -90 degrees, front is 0 degrees, and right is +90 degrees. Image elevation runs from +90 degrees at the top to -90 degrees at the bottom.

## 2. Release status and TODO

The calibration panoramas and videos are not part of the public release yet because some frames contain bystanders and must be privacy-filtered first. The data preparation and calibration scripts are also not released yet.

The cropped calibration examples below are included only to document the annotation and pose-estimation procedure.

## 3. Microphone calibration

We estimate microphone orientation and position separately. Orientation comes from AprilTag pose estimation. Position is reconstructed from two manual 2D annotations and the depth rendered for the particular scene reconstruction being evaluated. The figures below use `fountain-multi / mic_08` as an example.

### 3.1 Manual 2D position annotations

A sticker is placed on the ground directly below the microphone. In the input-camera calibration panorama, we manually annotate:

1. the center of that ground sticker (`mic_location_2d`); and
2. the physical center of the microphone (`mic_center_2d`).

![Fountain-multi mic 08 manual 2D annotations](example/fountain-multi_mic-08_2d-annotations.jpg)

The left panel is the unannotated crop. The right panel shows the microphone center in magenta and the ground sticker in cyan. The two pixels define rays from the input camera toward two vertically aligned points.

### 3.2 AprilTag orientation

An AprilTag rigidly attached to the microphone rig provides its orientation. We project the equirectangular calibration image into perspective views, detect the tag in a selected view, estimate its pose there, and transform that orientation back into the panorama world frame.

![Fountain-multi mic 08 AprilTag detection and pose](example/fountain-multi_mic-08_apriltag-pose.jpg)

The left panel shows the selected perspective view and detected tag boundary. The right panel is a closer view of the pose axes used for the rotation estimate. We retain only the resulting rotation in the released metadata; the raw AprilTag translation is discarded.

The stored `apriltag.rotation` is the rotation block `R` from `est_extrinsics_new`. It is a **world-to-camera (w2c)** rotation in the final microphone/object-camera convention, not a c2w rotation. A world-space point is transformed by

```text
p_mic = R_w2m p_world + t_w2m.
```

Here, “object frame” means the microphone-local frame after converting the raw AprilTag axes. The conventions are:

| Frame | Forward | Right | Up |
| --- | ---: | ---: | ---: |
| Microphone/object-camera (OpenCV) | `+Z` | `+X` | `-Y` |
| Panorama/world at the input-camera origin | `-X` | `+Y` | `+Z` |

Consequently, the microphone frame uses OpenCV camera axes (`+X` right, `+Y` down, `+Z` forward). In the panorama/world frame, `-X` is the input-camera forward direction, `+Y` is right, and `+Z` is up.

### 3.3 Reconstructing translation from depth

We do not publish a fixed 3D microphone location or translation. Instead, the evaluated reconstruction renders an equirectangular depth map at the input-camera pose, and the microphone center is reconstructed from that depth map plus the two released 2D points. This matters because different reconstruction runs can produce different depth at the same annotated pixel.

For an image of width `W` and height `H`, normalize a pixel as `u = x / W` and `v = y / H`. The calibration code converts it to polar angle `phi` and panorama angle `theta` using

```text
phi   = pi v
theta = 2 pi (1 - u)
ray(phi, theta) = [cos(theta) sin(phi), sin(theta) sin(phi), cos(phi)].
```

Let `(phi_g, theta_g)` denote the ground sticker, `(phi_c, theta_c)` the microphone center, and `d_g` the rendered depth sampled at the ground sticker. Assuming that the microphone center is vertically above the sticker, its ray distance is

```text
d_c = d_g sin(phi_g) / sin(phi_c),
C_world = d_c ray(phi_c, theta_c).
```

Once the run-specific microphone center `C_world` is available, the w2c translation is

```text
t_w2m = -R_w2m C_world.
```

This reconstruction keeps the released annotation independent of any one depth-rendering run while still producing a complete microphone extrinsic matrix for evaluation.

## Citation

If you use SonoWorld or the SonoScene360 dataset, please cite:

```bibtex
@article{jin2026sonoworld,
    title={SonoWorld: From One Image to a 3D Audio-Visual Scene},
    author={Jin, Derong and Chen, Xiyi and Lin, Ming C. and Gao, Ruohan},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2026}
}
```