config_improved.yaml

# Improved YOLOv8 Configuration for Higher Recall and mAP@0.5-0.95
# Optimized for Space Station Object Detection: FireExtinguisher, ToolBox, OxygenTank

# Dataset Configuration
path: ./  # Dataset root directory
train: data/train/images  # Train images
val: data/val/images      # Val images
test: data/test/images    # Test images

# Classes
nc: 3
names: ['FireExtinguisher', 'ToolBox', 'OxygenTank']

# Model Configuration - Use larger model for better performance
model: yolov8l.pt  # Upgraded from yolov8m.pt for better feature extraction
epochs: 50         # Increased from 20 for better convergence
batch_size: 8      # Reduced due to larger image size
imgsz: 832         # Increased from 640 for better small object detection
device: auto       # Use GPU if available

# Optimized Training Hyperparameters for Better Recall
lr0: 0.005         # Reduced initial learning rate for stability
lrf: 0.1           # Higher final LR ratio (was 0.01)
momentum: 0.937
weight_decay: 0.0005
warmup_epochs: 3.0
warmup_momentum: 0.8
warmup_bias_lr: 0.1

# Loss Function Tuning for Recall Improvement
box: 7.5           # Box loss gain
cls: 0.3           # Reduced class loss to emphasize detection over classification
dfl: 1.5           # DFL loss gain
label_smoothing: 0.1  # Added label smoothing for better generalization
nbs: 64
dropout: 0.1       # Added dropout for regularization

# Enhanced Data Augmentation for Better Recall and Generalization
hsv_h: 0.02        # Increased hue augmentation (was 0.015)
hsv_s: 0.8         # Increased saturation augmentation (was 0.7)
hsv_v: 0.5         # Increased value augmentation (was 0.4)
degrees: 10.0      # Added rotation augmentation (was 0.0)
translate: 0.2     # Increased translation (was 0.1)
scale: 0.9         # Increased scale augmentation (was 0.5)
shear: 2.0         # Added shear augmentation (was 0.0)
perspective: 0.0005 # Added perspective augmentation
flipud: 0.5        # Added vertical flip
fliplr: 0.5        # Horizontal flip probability
mosaic: 0.7        # Increased mosaic probability
mixup: 0.15        # Increased mixup for better generalization
copy_paste: 0.3    # Increased copy-paste augmentation
close_mosaic: 10   # Disable mosaic in last 10 epochs for precision

# Advanced Training Settings
patience: 15       # Early stopping patience
save_period: 5     # Save checkpoint every 5 epochs
val: True
plots: True
verbose: True

# Multi-scale Training for Better Detection at Different Scales
multiscale: True   # Enable multi-scale training
scale_range: [0.8, 1.2]  # Scale range for multi-scale training

# Test Time Augmentation (TTA) for inference
tta: True          # Enable TTA during validation

# Optimizer Settings
optimizer: AdamW   # AdamW optimizer for better convergence
cos_lr: True       # Cosine learning rate scheduler

# Advanced Augmentation Techniques
erasing: 0.4       # Random erasing probability
crop_fraction: 1.0 # Crop fraction for training

# Model EMA (Exponential Moving Average)
ema: True          # Enable model EMA for better validation performance
ema_decay: 0.9999  # EMA decay rate

# Additional Settings for Space Station Environment
# These settings are optimized for indoor/space station lighting conditions
auto_augment: randaugment  # Use RandAugment for automatic augmentation
augment: True      # Enable augmentation