Sound Source Localization Eval

---
name: sound-source-localization-eval
description: This evaluation probes a model's ability to spatially localize sound sources in images or video frames given an accompanying audio clip. It measures how accurately the predicted bounding box overlaps with ground-truth annotations provided by multiple human annotators. Use when the user wants to benchmark on Flickr SoundNet Testset, VGG-Sound Source (VGG-SS), or asks about evaluating this task. Reports cIoU.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2104.02691
  bibtex_key: chen2021localizingvisualsounds
  confidence: high
---

# sound-source-localization-eval

> Localizing Visual Sounds the Hard Way — Honglie Chen et al. (arXiv:2104.02691, 2021)

## What this evaluates

This evaluation probes a model's ability to spatially localize sound sources in images or video frames given an accompanying audio clip. It measures how accurately the predicted bounding box overlaps with ground-truth annotations provided by multiple human annotators.

## Datasets

- **Flickr SoundNet Testset** — total 250; splits: test (250)
- **VGG-Sound Source (VGG-SS)** — total ?; splits: test (-1)

## Metrics

- `cIoU` **(primary)** — range: [0, 1]
  - Consensus Intersection over Union (cIoU) measures the average overlap between predicted bounding boxes and consensus ground-truth boxes across multiple annotators.
- `AUC` — range: [0, 1]
  - Area Under Curve (AUC) summarizes localization performance across varying IoU thresholds or confidence scores.

## Input / output format

**Input**: A single image frame (224×224×3 tensor) paired with a 20-second audio clip (257×300 magnitude spectrogram).

**Output**: Bounding box coordinates predicting the sound source location in the image.

## Scoring recipe

```python
def compute_ciou(predictions, gold_boxes):
    total_ciou = 0.0
    for pred_box, annotator_boxes in zip(predictions, gold_boxes):
        ious = [calculate_iou(pred_box, g_box) for g_box in annotator_boxes]
        total_ciou += np.mean(ious)
    return total_ciou / len(predictions)
```

## Common pitfalls

- Training uses only center frames, but testing feeds the full-length audio spectrogram into the network.
- The test set size is fixed at 250 pairs, distinct from the variable training subsets (10k/144k).
- AUC implementation is not specified; readers must verify whether it uses IoU-threshold or confidence-score curves.

## Evidence (verbatim from paper)

> In order to quantitatively evaluate the proposed approach, we adopt the evaluation metrics used in [27, 31]: Consensus Intersection over Union (cIoU) and Area Under Curve (AUC) are reported for each model on two test sets, as detailed next.

## Citation

```bibtex
@misc{chen2021localizingvisualsounds,
  title={Localizing Visual Sounds the Hard Way},
  author={Honglie Chen et al.},
  year={2021},
  note={arXiv:2104.02691}
}
```

- arXiv: 2104.02691