🚀 YDS — YOLO Dataset Studio

Automatically build YOLO datasets from Twitch streams and videos

🇺🇸 English | 🇷🇺 Русский

🎯 What This Does

YDS downloads Twitch VODs or captures your screen, extracts frames, and uses an existing YOLO model to automatically generate bounding box labels. The result is a ready-to-use YOLO dataset (images + labels) for training.

Why? Creating YOLO datasets by hand is slow and repetitive. YDS automates most of this process while keeping you in control.

---

🚀 TL;DR — Quick Start

1. 📥 Capture screen or download Twitch VOD
2. 🤖 Auto-label frames with YOLO model
3. 🏷️ Manually verify & correct labels (add/remove/fix classes)
4. 📦 Split into train/val/test
5. ⚙️ Train YOLO model built-in

Note: Label verification is a manual step where you can add, remove, or fix bounding boxes and classes on every image.

---

🎮 Core Features

Feature	What It Does
📹 StreamCut	Download Twitch VODs, auto-extract & label frames (fully automated)
🎮 Screen Capture	Record gameplay + auto-label with YOLO (semi-automatic)
🖌️ Label Verification	Manual annotation tool — add/edit classes on screenshots
📦 Dataset Split	Auto partition train/val/test with custom ratios
⚙️ Training	Built-in YOLO training interface
📊 ONNX Benchmark	Test model inference speed & accuracy

---

🛠️ System Requirements

Component	Requirement
OS	Windows 10/11 (primary), Linux (manual setup)
Python	3.12+
GPU	Recommended for real-time capture/inference
CUDA	Tested with 11.8 & 12.8 (newer versions compatible)

---

📥 Installation & Run

Platform	Install	Run
Windows	setup.bat	run.bat
Linux	./setup.sh	./run.sh

---

🔄 Workflows

1️⃣ Screen Data Collection

1. Go to Dataset Tab → Click Start Data Collection
2. Configure: detection model, threshold, crop size, save interval
3. Frames & labels auto-saved to output_folder (see config.json)

---

2️⃣ Label Verification — Manual Correction

Review and correct AI-generated labels, or add new bounding boxes with assigned classes.

1. Go to Dataset Tab → Click Label Verification
2. Opens interactive OpenCV window
3. View auto-labeled frames (or unlabeled screenshots)
4. Add new bounding boxes with class assignment
5. Edit or delete existing boxes
6. Save annotations in YOLO format

⌨️ Keyboard Controls (click to expand)

Control	Action
🖱️ Right-Click	Add bounding box (current class)
🖱️ Left-Click	Remove box under cursor
🖱️ Drag Corner	Resize box
Dropdown	Change class
Spacebar	Toggle label visibility
N / P	Next / Previous image
D	Delete image
S	Save changes
Q	Quit

---

3️⃣ Split Dataset

Automatically splits your dataset into train / val / test folders using standard YOLO directory structure.

1. Go to Tools → Split Dataset
2. Select folder with images/ and labels/
3. Click Split → Uses fixed ratios: 70% train / 15% val / 15% test

Output:

dataset_split/
├── train/  (70%)
│   ├── images/
│   └── labels/
├── val/    (15%)
│   ├── images/
│   └── labels/
└── test/   (15%)
    ├── images/
    └── labels/

---

4️⃣ Training

1. Go to Training Tab
2. Select data.yaml
3. Set basic parameters: epochs, image size, batch size
4. (Optional) Expand Advanced Options for augmentation & performance tuning
5. Click Start Training

⚡ Advanced Options (click to expand)

Augmentation Settings:

{
  "mosaic": 1.0,        // Multi-scale training
  "mixup": 0.1,         // Image mixing
  "fliplr": 0.5,        // Horizontal flip
  "flipud": 0.0,        // Vertical flip
  "scale": 0.5,         // Random scaling
  "hsv_h": 0.015,       // Hue shift
  "hsv_s": 0.7,         // Saturation
  "hsv_v": 0.4          // Brightness
}

Regularization & Performance:

{
  "amp": true,          // Automatic Mixed Precision (faster training)
  "patience": 20,       // Early stopping threshold
  "save_period": 10,    // Save checkpoint every N epochs
  "weight_decay": 0.0005
}

Training Output Structure:

runs/valorant/exp1/
├── weights/
│   ├── best.pt          # Best model (highest mAP)
│   └── last.pt          # Last epoch
├── results.csv          # All metrics
└── plots/
    ├── confusion_matrix.png
    ├── F1_curve.png
    ├── PR_curve.png
    └── results.png

---

5️⃣ StreamCut & Auto-Labeling

What it does: Download Twitch VODs, automatically extract frames, run YOLO inference, save labeled dataset.

1. Go to Tools → Open StreamCut
2. Add Twitch VOD URLs
3. Choose: download quality, YOLO model, detection threshold, worker threads
4. Downloads to stream/raw_streams, outputs labeled frames to stream/dataset

How it works:

• You provide a pre-trained YOLO model
• Tool downloads VOD and splits into segments
• Runs inference on every N frames
• Saves detections as YOLO-format labels
• Result: fully labeled dataset ready for training
• Button "Sync" отмечает уже скачанные стримы

Key: Better input model = better auto-labeled dataset

StreamCut Interface:

---

🤖 ONNX Model Benchmarking

What it does: Run inference on multiple ONNX models to compare detection results and model outputs.

1. Go to Tools → Benchmark ONNX Models
2. Select folder with .onnx model files
3. Click Run Benchmark → Runs inference on each model with test images
4. Results show per-model comparison table:

Model                                    Shape        nCls  #   Classes      Conf1 
─────────────────────────────────────────────────────────────────────────────────
1.5kR6.onnx                              [8400, 6]    1     45  0            0.92
CS2 by Kwnema.onnx                       [8400, 7]    3     12  0,2,3        0.85
ABI_v3.onnx                              [8400, 84]   80    28  5,16,24      0.78

Why use it: Compare how different model architectures and quantizations detect objects in your test images.

---

📥 Model Manager

What it does: One-click download of official YOLO weights for quick experimentation.

1. Go to Tools → YOLO Model Download
2. Select model size: N (nano) / S (small) / M (medium) / L (large) / X (xlarge)
3. Select version: v8 / v10 / v11 / v12
4. Click Download → Saves to models/ folder

Why use it: Start training without hunting for model files. Pre-trained weights ready to use immediately.

---

⚙️ Configuration

• configs/config.json — GUI defaults, dataset collection, label verification settings
• configs/configStreamCut.json — StreamCut settings (download workers, thresholds, etc.)

Full reference: docs/HELP.md

---

📁 Project Structure

YolovTrainGui/
│
├── 📄 GUI.py                          # Main application
├── 📄 setup.bat / run.bat             # Windows scripts
├── 📄 setup.sh / run.sh               # Linux scripts
├── 📋 requirements.txt                # Python dependencies
├── 📖 README.md                       # English guide (you are here)
│
├── 📂 docs/
│   ├── 📖 README_ru.md                # Russian guide
│   ├── 📖 HELP.md                     # Configuration reference (English)
│   └── 📖 HELP_ru.md                  # Configuration reference (Russian)
│
├── 📂 Core/                           # Core modules
│   ├── 🚀 train.py                    # YOLO training pipeline
│   ├── 🎬 StreamCut.py                # VOD processor
│   ├── 🎮 semiauto_dataset_collector.py  # Screen capture
│   ├── 🖌️ labelConfig.py              # Label verification tool
│   └── ✂️ splitDatasetFiles.py        # Dataset splitter
│
├── 📂 configs/                        # Configuration files
│   ├── ⚙️ config.json                 # Main settings
│   └── 🎬 configStreamCut.json        # StreamCut settings
│
├── 📂 benchmark/                      # Performance testing
│   ├── 📊 benchmark.py                # ONNX benchmarking
│   ├── infer_function.py
│   └── 📁 models/                     # ONNX models
│
├── 📂 models/                         # YOLO weights (.pt files)
├── 📂 datasets/                       # Training datasets
│
├── 📂 stream/                         # StreamCut output
│   ├── 📁 raw_streams/                # Downloaded VODs
│   ├── 📁 chunks/                     # Split segments
│   └── 📁 dataset/                    # Labeled output
│
├── 📂 runs/                           # Training results
│   └── 📁 detect/train*/              # Model checkpoints & metrics
│
├── 📂 docs/
│   └── 📁 images/                     # Screenshots & diagrams
│
└── 📂 utils/                          # Utilities
    └── 📁 ffmpeg/                     # FFmpeg binaries

---

❓ FAQ

Q: Can I use my own YOLO model?
A: Yes, provide any .pt model file.

Q: Does it work without GPU?
A: Yes, but screen capture and VOD processing will be much slower.

Q: Can I combine datasets later?
A: Yes, all datasets are in standard YOLO format.

Q: What if I don't have a YOLO model yet?
A: Use Model Manager to download v8/v12 presets, or train a basic one first.

Q: I stopped StreamCut mid‑process. What should I do?
A: Delete stream/resume.json and желательно очистить stream/dataset/ if you plan to re‑run.
Resume file tracks completed chunks; if you interrupted, it may mark them as finished.

---

👥 Who This Is For

YDS is a complete GUI solution for building YOLO datasets without CLI scripts.

Use YDS if you:

• ✅ Want to build YOLO datasets without manual labeling
• ✅ Work with Twitch streams or gameplay videos
• ✅ Need fast iteration on object detection models
• ✅ Want all tools in one place (capture, label, split, train) instead of gluing CLI scripts together
• ✅ Prefer GUI over command line for dataset management

---

🛣️ Roadmap (Ideas)

• YouTube support for StreamCut
• Custom dataset merge tools
• Linux GUI improvements
• Multi-monitor capture

---

🔗 Quick Links

Resource	Description
📖 Configuration Reference	Detailed config.json & StreamCut settings
📖 Russian Guide	🇷🇺 Полное руководство на русском
⚙️ Main Config	GUI defaults & dataset settings
🎬 StreamCut Config	VOD download & processing settings
📄 License	MIT License

---

Made with ❤️ for the computer vision community

⭐ Star on GitHub | 🐛 Report Issue

Happy detecting! 🚀