AutoDetect — an AutoML framework for object detection that automatically trains ensemble YOLO models, optimizes hyperparameters with Optuna, and fuses predictions using Weighted Box Fusion (WBF) for state-of-the-art results.

Example Notebooks: simple Notebook, list datasets Notebook

Video:

Features:

1. Ensemble Learning: Combines predictions from multiple YOLO architectures (yolov8, yolo11, yolo12) with different input resolutions.

2. Optuna Hyperparameter Search: Automatically finds optimal conf_thresh, iou_thresh, and box filtering parameters for WBF.

3. Weighted Box Fusion (WBF): Advanced box merging strategy (avg or max confidence aggregation) for robust inference.

4. GPU-Accelerated: Built on PyTorch and Ultralytics for fast training and inference.

5. Stratified Validation Sampling: Configurable validation subset for efficient hyperparameter tuning.

6. Reproducibility: Full seed control and optional deterministic mode for consistent results.

Installation:

pip install git+https://github.com/Antonoof/AutoDetect.git@main

🗂️ Dataset Structure:

dataset/
├── train/
│   ├── images/   # Training images (.jpg, .png, etc.)
│   └── labels/   # YOLO-format .txt labels (class_id x_center y_center w h)
├── val/
│   ├── images/   # Validation images
│   └── labels/   # Corresponding validation labels
└── test/
    └── images/   # Test images for inference (no labels required)

Training: AutoDetect

from autodetect import AutoDetect

ad = AutoDetect(
    train="path/to/train",          # Path to train folder
    val="path/to/val",              # Path to val folder
    model='x',                      # Model size modifier: 'n', 's', 'm', 'l', 'x'
    epochs=30,                      # Number of training epochs
    model_config=(                  # Ensemble: (model_type, img_size)
        ('yolov8', 1024),
        ('yolo11', 960),
        ('yolo12', 768),
    ),
    device=None,                    # 'cpu', 0 or [0, 1] - two GPU
    warmup=False,                   # Enable RandomSearch to find best params YOLO
    inference_speed=-1,             # Target FPS for model selection (-1 = best accuracy)
    seed=42                         # Random seed for reproducibility
)

ad.fit()  # Start training and ensemble building

Inference: ADPredict

from autodetect import ADPredict

predictor = ADPredict(
    image_paths=test_images,               # test dir (images in test dir)
    val_images_path=val_data,              # Validation images for Optuna tuning
    models_dir="result",                   # Directory with trained models default="result"
    output_dir="predictions",              # Where to save results
    device='auto',                         # 0, 'cpu', or 'auto'
    seed=42,
    deterministic=False,                   # Enable deterministic algorithms
    optuna_trials=30,                      # Number of Optuna hyperparameter trials
    val_samples=200,                       # Subset of val images for tuning
    wbf_conf_type="avg",                   # WBF confidence: 'avg' or 'max'
    conf_range=(0.005, 0.5),               # Search range for confidence threshold
    iou_range=(0.3, 0.7),                  # Search range for IoU threshold
    skip_box_range=(0.001, 0.05),          # Search range for box filtering
    model_weights=None                     # Optional: custom model weights example [0.9, 0.8, 1.3] for 3 models
)

# Run prediction
results = predictor.predict()

Output Format

Predictions are saved in output_dir/ (default: "predictions") as <image_name>.txt files.

File Format

Each line represents one detected object in space-separated format:

Field	Type	Range	Description
class_id	int	0, 1, 2, ...	Predicted class index
confidence	float	[0.0, 1.0]	Confidence score of the prediction
x_center	float	[0.0, 1.0]	Normalized X coordinate of box center
y_center	float	[0.0, 1.0]	Normalized Y coordinate of box center
width	float	[0.0, 1.0]	Normalized box width
height	float	[0.0, 1.0]	Normalized box height

Example <image_name>.txt

0 0.625189 0.497187 0.149444 0.194430 0.050108
0 0.596752 0.961163 0.929023 0.077674 0.140234

Roadmap updates:

Version	Planned Improvements
v0.2.*	Enhanced hyperparameter stability, improved training convergence, better logging
v0.3.*	GPU memory optimization, automatic model sizing based on image count/resolution, inference timing metrics
v0.4.*	LR warmup scheduler, DINO-based validation split for stable training, time-aware hyperparameter search
v0.5.*	Integration with 3LC or similar for active learning and bbox refinement
v0.6.*	Advanced augmentation pipeline, use Duality synthetic image generation during training
v0.7.*	SAHI integration for small-object detection, sliding-window inference support