README.md

---
license: mit
tags:
  - object-detection
  - apex-legends
  - yolov8
  - ultralytics
---

# apex_enemy_detect

> [Русская версия](README-RU.md)

YOLOv8 models for detecting enemies in Apex Legends gameplay footage. Two variants: nano (speed) and medium (accuracy).

[Apex Enemy Detect Demo](https://github.com/PSImera/apex_enemy_detect_demo) — a tool to analyze gameplay videos, detect enemies, and fix stretched resolution issues.

## Models

| | **YOLOv8n (nano)** | **YOLOv8m (medium)** |
|---|---|---|
| File | `apex_detect_v8n_v2.1.pt` | `apex_detect_v8m_v2.1.pt` |
| Parameters | ~3.2M | ~25.9M |
| Precision | 0.932 | 0.943 |
| Recall | 0.877 | 0.883 |
| mAP@50 | 0.930 | 0.938 |
| mAP@50-95 | 0.756 | 0.796 |
| Better for | Speed / low-end GPU | Accuracy |

## Usage

```python
from ultralytics import YOLO

model = YOLO("apex_detect_v8m_v2.1.pt")
results = model.predict("frame.jpg", conf=0.35, iou=0.5, imgsz=640)
```

Or use it automatically via the [Apex Enemy Detect Demo](https://github.com/PSImera/apex_enemy_detect_demo) app — the models are loaded from here on first run.

## Training Setup

| Parameter | Value |
|---|---|
| Dataset | `apex_detect_v2_p1_converted` |
| Epochs | 200 (patience 100) |
| Batch size | 16 |
| Image size | 640×640 |
| Optimizer | AdamW (lr=0.001) |
| Augmentations | HSV (S/V ±0.3), horizontal flip (p=0.5), random erasing (p=0.4) |

Fine-tuned from Ultralytics COCO pretrained weights on a custom Apex Legends enemy dataset.

## Training Curves

<table>
<tr><th>YOLOv8n (nano)</th><th>YOLOv8m (medium)</th></tr>
<tr>
<td><img src="run_v8n_v2.1/results.png"></td>
<td><img src="run_v8m_v2.1/results.png"></td>
</tr>
</table>

## Precision-Recall Curves

<table>
<tr><th>YOLOv8n (nano)</th><th>YOLOv8m (medium)</th></tr>
<tr>
<td><img src="run_v8n_v2.1/BoxPR_curve.png"></td>
<td><img src="run_v8m_v2.1/BoxPR_curve.png"></td>
</tr>
</table>