---
license: agpl-3.0
metrics:
  - map50
  - map50-95
  - precision
  - recall
  - f1
base_model:
- Ultralytics/YOLO26
pipeline_tag: object-detection
tags:
- yolo
- vision
- object-detection
- vrchat
- vrchat-avatar
- vrchat-nametag
- avatar-detection
model-index:
  - name: VRChat_Avatar_Detection (YOLO26m)
    results:
      - task:
          type: object-detection
          name: Object Detection
        metrics:
          - name: mAP@50
            type: map50
            value: 96.30
          - name: mAP@50-95
            type: map50-95
            value: 81.30
          - name: Precision
            type: precision
            value: 96.50
          - name: Recall
            type: recall
            value: 92.80
          - name: F1-Score
            type: f1
            value: 94.61
---
# VRChat Avatar Detection

<p style="display: grid;grid-template-columns: auto auto">
  <a href="https://huggingface.co/Gureumi/VRChat_Avatar_Detection/blob/main/result_1920.jpg">
    <img src="result_1920.jpg" alt="1920x1080 image">
  </a>
  <a href="https://huggingface.co/Gureumi/VRChat_Avatar_Detection/blob/main/result_7680.jpg">
    <img src="result_7680.jpg" alt="7680x4320 image">
  </a>
  <a href="https://huggingface.co/Gureumi/VRChat_Avatar_Detection/blob/main/result_avatar&amp;nametag.png">
    <img src="result_avatar&amp;nametag.png" alt="avatar&nametag detection image">
  </a>
  <a href="https://huggingface.co/Gureumi/VRChat_Avatar_Detection/blob/main/result_avatar&amp;nametag_2.png">
    <img src="result_avatar&amp;nametag_2.png" alt="avatar&nametag detection image">
  </a>
</p>

## Model Description

This model is the YOLO object detection model used in programs that detect avatars and name tags in a VRChat environment to determine which avatar belongs to whom.

## Dataset Information

- **Train Images**: 13,565
- **Validation Images**: 1,507
- **Image Size**: 1920x1080 ~ 15360x8360
- **Data Source**: Photos captured using the in-game camera over 7 years of VRChat gameplay, along with frame data extracted from 1,707 hours of recorded video footage.

### Data Filtering & Preprocessing
- **Included**: Humanoid 3D avatars only.
- **Excluded**: Non-humanoid avatars, Fallback avatars, default robot avatar, and Impostor avatars.

## Training Configuration
- **Task**: Object Detection
- **Base Framework**: Ultralytics YOLO26
- **Input Resolution(imgsz)**: 1920

## Usage

```python
from ultralytics import YOLO

model = YOLO("yolo26m_vrchat_avatar.pt")
results = model.predict(source="vrchat_screenshot.png", imgsz=1920, conf=0.5)
results[0].show()
```

## Performance

| Metric     | Score (%)       |
| ---------- | --------------- |
| mAP@50     | 96.30           |
| mAP@50-95  | 81.30           |
| Precision  | 96.50           |
| Recall     | 92.80           |
| F1 Score   | 94.61           |
| Parameters | 20.35M          |
| FLOPs      | 68.1B           |
| Inference  | 4.7ms (NVIDIA H100)<br>12.2ms (RTX 5070Ti) |

## Per-Class Performance

| Class   | Images | Instances | Precision (%) | Recall (%) | mAP@50 (%) | mAP@50-95 (%) |
| ------- | ------ | --------- | ------------- | ---------- | ---------- | ------------- |
| avatar  | 1455   | 2777      | 95.90         | 90.80      | 95.80      | 79.80         |
| nametag | 121    | 268       | 97.10         | 94.80      | 96.80      | 82.80         |
| all     | 1507   | 3045      | 96.50         | 92.80      | 96.30      | 81.30         |

## Known Limitations

Performance may degrade in:

* Severe motion blur or particle & SFX hell environments
* Humanoid furry (I don't have many furry humanoid photos)
* Types of humanoid avatars I have never seen