🧠 CS2 CT/TERO Detection

Lightweight AI project to detect and classify Terrorists and Counter-Terrorists in Counter-Strike 2 gameplay, using custom-trained YOLOv5 on real CS2 footage.

---

🎯 Project Overview

This project performs object detection on Counter-Strike 2 gameplay footage to identify players as either:

• 🔵 Terrorists (T)
• 🔶 Counter-Terrorists (CT)

It leverages a yolo11 model trained on real game data collected and labeled from the current map pool. The detector processes videos and returns a labeled version with bounding boxes around each identified player.

---

🚀 Features

• 🎥 Video-based object detection (offline inference)
• 🧠 Custom yolo11 model trained on labeled gameplay footage
• 📸 Generates labeled video frames or full videos with bounding boxes
• 📦 Deployable as a local script (CLI)

---

🗂 Dataset Preparation Workflow

🎮 Map Pool

The dataset was built using replays from the official CS2 map pool as of early 2025:

• Dust2
• Mirage
• Nuke
• Ancient
• Anubis
• Inferno
• Train

🎥 Frame Extraction

Two gameplays per map were used (14 demos total). From each, frames were extracted at a rate of 1 frame per minute using ffmpeg:

ffmpeg -i match_demo.mp4 -vf fps=1 frames/mapname_%04d.png

This generated approximately 2000 frames per video.

🚩 Frame Filtering

To remove uninformative frames (scoreboards, blank screens, etc.), a custom filtering script was written. Final dataset: ~2400 high-quality frames , down from ~30,000 raw frames.

Script available here: tools/filter_frames.py

🏷️ Annotation Process

• Used LabelStudio (labelstud.io) for manual labeling
• Leveraged LabelStudio ML Backend (GitHub) with a model pre-trained on 100 images to assist annotation
• Labels: CT, T

---

🧠 Model Training

• Model: yolo11s (Ultralytics)
• Epochs: 100
• Input Size: 640x640
• Framework: Google Colab

Colab notebook used for training: EdjeElectronics YOLO Colab

📊 Training Metrics

The training curves indicate:

• 📉 Losses (box, cls, DFL) consistently decrease across training and validation.
• ✅ Precision and recall stabilize around ~0.90 after 50 epochs.
• 📈 Final mAP50 peaks near 0.92 , while mAP50–95 reaches ~0.71 , indicating reliable bounding box accuracy and class separation.

These results are encouraging for a lightweight, fast-inference model trained on a small, curated dataset.

---

🛠️ Usage

1. Clone the repository

git clone https://github.com/moijesuis2enmoi/cs2-ct-tero-detector.git
cd cs2-ct-tero-detector

2. Install dependencies

pip install -r requirements.txt

If you have an NVIDIA GPU, install the GPU-enabled version of PyTorch:

pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

3. Run detection on a video file

python detect.py --source path/to/cs2_video.mp4 --weights model.pt

The output will be saved in the runs/detect/ directory.

📎 This is a lightweight prototype designed for portfolio/demo purposes. Accuracy may vary by map, skin, lighting, and POV. Detection is limited to what the dataset covers. Potential for extension into real-time systems with tracking, webcam input, or stream overlays.

---

🧪 Real-Time Detection (Optional)

Once the model is trained (best.pt), you can run live inference with:

python detect.py --model runs/detect/train/weights/best.pt --source screen1 --resolution 1280x720

Available options:

Argument	Description
--model	Path to trained YOLO model (.pt file)
--source	Input source: image, folder, video,usb0,picamera0, or screen1
--thresh	Confidence threshold (default:0.5)
--resolution	Force video stream resolution (e.g.,1280x720)
--record	Record output video (requires a resolution to be set)

Runtime Controls

• Q: Quit
• S: Pause
• P: Save a screenshot (capture.png)

---

📸 Examples

Here are some example outputs from the detector:

---

📂 Project Structure

├── model.pt               # Trained yolo11 weights
├── detect.py              # Main inference script
├── train/                 # Training results, weights, and logs
├── tools/
│   └── filter_frames.py   # Frame filtering utility
├── assets/                # Screenshots and preview GIFs
├── requirements.txt       # Dependencies
└── README.md

---

👤 Author

Project by moijesuis2enmoi

Feel free to fork, suggest improvements, or contact for collaboration opportunities.