AVATAR: What’s Making That Sound Right Now? Video-centric Audio-Visual Localization

AVATAR stands for Audio-Visual localizAtion benchmark for a spatio-TemporAl peRspective in video.

AVATAR is a benchmark dataset designed to evaluate video-centric audio-visual localization (AVL) in complex and dynamic real-world scenarios.
Unlike previous benchmarks that rely on static image-level annotations and assume simplified conditions, AVATAR offers high-resolution temporal annotations over entire videos. It supports four challenging evaluation settings:
Single-sound, Mixed-sound, Multi-entity, and Off-screen.

📄 Paper (ICCV 2025)
🌐 Project Website
📁 Code & Data Viewer

---

📦 Dataset Structure

The dataset consists of the following files:

File	Description
video.zip	~3.8GB of .mp4 video clips
metadata.zip	~1.6GB of annotations (bounding boxes, segmentation masks, scenario tags)
vggsound_10k.txt	List of 10,000 training video IDs from VGGSound
code/	AVATAR benchmark evaluation code

Each annotated frame includes:

• Visual bounding boxes and segmentation masks for sound-emitting objects
• Audio-visual category labels aligned to the active sound source at each timestamp
• Instance-level scenario labels (e.g., Off-screen, Mixed-sound)

---

📊 Dataset Statistics

AVATAR provides detailed quantitative statistics to help users understand its scale and diversity.

Type	Count
Videos	5,000
Frames	24,266
Off-screen	670

Scenario Type	Instances
Total	28,516
Single-sound	15,372
Multi-entity	9,322
Mixed-sound	3,822

---

🧪 Scenarios and Tasks

AVATAR supports fine-grained scenario-wise evaluation of AVL models:

1. Single-sound: One sound-emitting instance per frame
2. Mixed-sound: Multiple overlapping sound sources (same or different categories)
3. Multi-entity: One sounding instance among multiple visually similar ones
4. Off-screen: No visible sound source within the frame

🔍 You can evaluate your model using:

• Consensus IoU (CIoU)
• AUC
• Pixel-level TN% (for Off-screen)

---

🧩 Audio-Visual Category Diversity

AVATAR spans 80 audio-visual categories covering a wide range of everyday domains, including:

• Human activities (e.g., talking, singing)
• Music performances (e.g., violin, drum, piano)
• Animal sounds (e.g., dog barking, bird chirping)
• Vehicles (e.g., car engine, helicopter)
• Tools and machines (e.g., chainsaw, blender)

Such diversity enables a comprehensive evaluation of model generalizability across varied audio-visual contexts.

---

📝 Example Metadata Format

{
  "video_id": str,
  "frame_number": int,
  "annotations": [
    { // instance 1 (e.g., man)
      "segmentation": [ // (x, y) annotated RLE format
            [float, float], 
            ...
          ],
      "bbox": [float, float, float, float], // (l, t, w, h),
      "scenario": str, // "Single-sound", "Mixed-sound", "Multi-entity", "Off-screen"
      "audio_visual_category": str,
    },
    { // instance 2 (e.g., piano)
      ...
    }, 
    ...
  ]
}