inference.py

"""
Wildlife Detection with YOLOv26 — Inference Script
===================================================
Supports RGB and thermal drone imagery.

Usage:
    python inference.py --model rgb --source path/to/image.jpg
    python inference.py --model thermal_merged --source path/to/thermal/ --save
    python inference.py --model matched_rgb --source image.jpg --conf 0.3 --show

Available models:
    thermal_original  — Baseline thermal model
    thermal_merged    — Refined thermal model (more training data)
    rgb               — Primary RGB model
    matched_rgb       — RGB model trained on matched RGB/thermal pairs
    matched_thermal   — Thermal model trained on matched RGB/thermal pairs
"""

import argparse
from pathlib import Path
from ultralytics import YOLO


MODELS = {
    "thermal_original": "thermal_original/weights/best.pt",
    "thermal_merged":   "thermal_merged/weights/best.pt",
    "rgb":              "rgb/weights/best.pt",
    "matched_rgb":      "matched_rgb/weights/best.pt",
    "matched_thermal":  "matched_thermal/weights/best.pt",
}


def load_model(name: str) -> YOLO:
    """Load a model by name or direct path."""
    path = MODELS.get(name, name)
    print(f"Loading model: {path}")
    return YOLO(path)


def run_inference(
    model_name: str = "rgb",
    source: str = "0",
    imgsz: int = 1024,
    conf: float = 0.25,
    iou: float = 0.45,
    show: bool = False,
    save: bool = False,
    save_txt: bool = False,
    project: str = "detections",
    name: str = "predict",
    device: str = "",
):
    """Run inference and return results."""
    model = load_model(model_name)

    results = model.predict(
        source=source,
        imgsz=imgsz,
        conf=conf,
        iou=iou,
        show=show,
        save=save,
        save_txt=save_txt,
        project=project,
        name=name,
        device=device if device else None,
    )

    for i, result in enumerate(results):
        n = len(result.boxes)
        print(f"[Image {i+1}] {n} detection(s)")
        for box in result.boxes:
            cls_id = int(box.cls.item())
            cls_name = result.names[cls_id]
            conf_val = box.conf.item()
            xyxy = [round(v, 1) for v in box.xyxy[0].tolist()]
            print(f"  {cls_name:15s}  conf={conf_val:.2f}  bbox={xyxy}")

    return results


def compare_modalities(
    rgb_source: str,
    thermal_source: str,
    conf: float = 0.25,
    imgsz: int = 1024,
):
    """
    Compare RGB vs thermal detections on co-registered image pairs.
    Useful for the matched dataset experiments.
    """
    rgb_model     = load_model("matched_rgb")
    thermal_model = load_model("matched_thermal")

    rgb_results     = rgb_model.predict(rgb_source,     imgsz=imgsz, conf=conf, verbose=False)
    thermal_results = thermal_model.predict(thermal_source, imgsz=imgsz, conf=conf, verbose=False)

    for i, (r_rgb, r_thm) in enumerate(zip(rgb_results, thermal_results)):
        print(f"\n--- Pair {i+1} ---")
        print(f"  RGB     detections: {len(r_rgb.boxes)}")
        print(f"  Thermal detections: {len(r_thm.boxes)}")

    return rgb_results, thermal_results


def main():
    parser = argparse.ArgumentParser(description="Wildlife YOLOv26 Inference")
    parser.add_argument(
        "--model",
        default="rgb",
        choices=list(MODELS.keys()) + ["custom"],
        help="Model to use. Pass a file path with --model custom --weights <path>.",
    )
    parser.add_argument("--weights", default=None, help="Direct path to .pt weights file.")
    parser.add_argument("--source", default="0", help="Image/video/folder path or webcam index.")
    parser.add_argument("--imgsz", type=int, default=1024, help="Inference image size.")
    parser.add_argument("--conf", type=float, default=0.25, help="Confidence threshold.")
    parser.add_argument("--iou", type=float, default=0.45, help="NMS IoU threshold.")
    parser.add_argument("--show", action="store_true", help="Display results.")
    parser.add_argument("--save", action="store_true", help="Save annotated images.")
    parser.add_argument("--save-txt", action="store_true", help="Save YOLO-format labels.")
    parser.add_argument("--project", default="detections", help="Output project folder.")
    parser.add_argument("--name", default="predict", help="Output run name.")
    parser.add_argument("--device", default="", help="CUDA device, e.g. '0' or 'cpu'.")
    parser.add_argument(
        "--compare",
        nargs=2,
        metavar=("RGB_SOURCE", "THERMAL_SOURCE"),
        help="Compare RGB and thermal models on co-registered pairs.",
    )
    args = parser.parse_args()

    if args.compare:
        compare_modalities(
            rgb_source=args.compare[0],
            thermal_source=args.compare[1],
            conf=args.conf,
            imgsz=args.imgsz,
        )
        return

    model_name = args.weights if (args.model == "custom" and args.weights) else args.model

    run_inference(
        model_name=model_name,
        source=args.source,
        imgsz=args.imgsz,
        conf=args.conf,
        iou=args.iou,
        show=args.show,
        save=args.save,
        save_txt=args.save_txt,
        project=args.project,
        name=args.name,
        device=args.device,
    )


if __name__ == "__main__":
    main()