config/default.yaml

# =============================================================================
# Wildlife Detector — Malilangwe Trust
# Master configuration file
# =============================================================================
# All paths are relative to the project root unless marked absolute.
# Override per-environment with config/local.yaml (git-ignored).

# -- Project metadata --------------------------------------------------------
project:
  name: "Malilangwe Wildlife Detector"
  version: "0.1.0"
  description: >
    Aerial drone wildlife detection and tracking for the Malilangwe Trust,
    Zimbabwe. Uses YOLOv11 fine-tuned on the WAID dataset with BoT-SORT
    multi-object tracking.

# -- Paths -------------------------------------------------------------------
paths:
  # WAID dataset root (contains images/ and labels/ with train/valid/test splits)
  dataset_root: "WAID/WAID"
  # Merged dataset root (Phase A+ — generated by scripts/merge_datasets.py)
  merged_dataset_root: "data/merged"
  # Directory for trained model weights
  weights_dir: "weights"
  # Default model checkpoint for inference.
  default_model: "weights/best.pt"
  # Where inference results (images/videos) are saved
  output_dir: "outputs"
  # Temporary / scratch directory
  temp_dir: "tmp"

# -- Dataset -----------------------------------------------------------------
# WAID (Wildlife Aerial Images from Drone)
# Source: https://github.com/xiaohuicui/WAID
# Classes defined in WAID/WAID/classes.txt — order matches label indices 0-5.
dataset:
  name: "WAID"
  num_classes: 6
  class_names:
    - "sheep"
    - "cattle"
    - "seal"
    - "camelus"
    - "kiang"
    - "zebra"
  # Dataset is pre-split into train/valid/test subdirectories
  split_dirs:
    train: "train"
    val:   "valid"
    test:  "test"
  seed: 42

# -- Detection ---------------------------------------------------------------
detection:
  # Model variant: yolo11n / yolo11s / yolo11m / yolo11l / yolo11x
  model_variant: "yolo11n"
  # Confidence threshold for detections
  confidence_threshold: 0.25
  # IoU threshold for NMS
  iou_threshold: 0.45
  # Max detections per image
  max_detections: 100
  # Input image size (pixels, square)
  image_size: 640
  # Device: "cpu", "cuda", "cuda:0", "mps" (Apple Silicon)
  device: "cpu"
  # Half-precision inference (requires GPU)
  half_precision: false
  # Augment inference (TTA — test-time augmentation)
  augment: false

# -- Training ----------------------------------------------------------------
training:
  epochs: 100
  batch_size: 16
  image_size: 640
  optimizer: "AdamW"
  learning_rate: 0.001
  weight_decay: 0.0005
  patience: 15          # Early-stopping patience (epochs)
  # Resume from checkpoint (path or false)
  resume: false
  # Class weighting — addresses severe imbalance in WAID dataset
  # (sheep: 91k instances vs kiang: 3k). Higher weight = more focus.
  # Set null to disable. Weights are inverse-frequency-based.
  class_weights:
    - 0.2    # sheep   (91,496 — overrepresented)
    - 0.4    # cattle  (44,245)
    - 1.0    # seal    (15,762)
    - 3.0    # camelus  (4,676 — underrepresented)
    - 4.0    # kiang    (3,312 — most underrepresented)
    - 3.5    # zebra    (3,792 — underrepresented)
  # Augmentation toggles
  augmentation:
    hsv_h: 0.015
    hsv_s: 0.7
    hsv_v: 0.4
    flipud: 0.5          # Aerial images benefit from vertical flip
    fliplr: 0.5
    mosaic: 1.0
    mixup: 0.1
    scale: 0.5

# -- Tracking (BoT-SORT) ----------------------------------------------------
tracking:
  tracker: "botsort"
  # Config file shipped with ultralytics (or custom path)
  tracker_config: "botsort.yaml"
  # Re-identification model (set null to disable ReID)
  reid_model: null
  # Track buffer — frames to keep lost tracks alive
  track_buffer: 30
  # Minimum track length to report (frames)
  min_track_length: 3

# -- Visualization -----------------------------------------------------------
visualization:
  # Bounding-box line thickness (pixels)
  line_thickness: 2
  # Font scale for labels
  font_scale: 0.6
  # Show confidence scores on boxes
  show_confidence: true
  # Show track IDs (only meaningful in video/tracking mode)
  show_track_id: true
  # Colour palette per class (BGR format for OpenCV)
  class_colors:
    sheep:    [0, 200, 0]        # green
    cattle:   [0, 165, 255]      # orange
    seal:     [255, 200, 0]      # cyan-blue
    camelus:  [0, 215, 255]      # gold
    kiang:    [180, 105, 255]    # pink
    zebra:    [255, 255, 255]    # white

# -- Logging -----------------------------------------------------------------
logging:
  level: "INFO"           # DEBUG | INFO | WARNING | ERROR
  log_to_file: true
  log_dir: "logs"

# -- Edge deployment (future) ------------------------------------------------
edge:
  # Target: "jetson_nano" | "jetson_orin" | "raspberry_pi_5" | "desktop"
  target_device: "desktop"
  # Export format: "torchscript" | "onnx" | "engine" (TensorRT)
  export_format: "onnx"
  # INT8 quantization (requires calibration dataset)
  quantize: false