zjsh_yolov11/推理图片反向上传CVAT/obb/tuili_save_txt_f.py

import os
import cv2
import numpy as np
from pathlib import Path
from ultralytics import YOLO

IMG_EXTENSIONS = {'.jpg', '.jpeg', '.png', '.bmp', '.tif', '.tiff', '.webp'}


def rotate_box_to_corners(cx, cy, w, h, angle_rad):
    corners = np.array([
        [-w / 2.0, -h / 2.0],
        [ w / 2.0, -h / 2.0],
        [ w / 2.0,  h / 2.0],
        [-w / 2.0,  h / 2.0]
    ], dtype=np.float32)

    cos_a = np.cos(angle_rad)
    sin_a = np.sin(angle_rad)
    R = np.array([[cos_a, -sin_a],
                  [sin_a,  cos_a]], dtype=np.float32)

    rotated = corners.dot(R.T)
    rotated += np.array([cx, cy], dtype=np.float32)
    return rotated


def reorder_points_fixed(pts):
    pts = np.array(pts)
    pts_sorted_y = pts[np.argsort(pts[:, 1])]
    top_two = pts_sorted_y[:2]
    bottom_two = pts_sorted_y[2:]
    top_left = top_two[np.argmin(top_two[:, 0])]
    top_right = top_two[np.argmax(top_two[:, 0])]
    bottom_left = bottom_two[np.argmin(bottom_two[:, 0])]
    bottom_right = bottom_two[np.argmax(bottom_two[:, 0])]
    return [top_left, top_right, bottom_right, bottom_left]


def process_obb_and_save_yolo_txt(model_path, image_dir, output_dir="./inference_results", conf_thresh=0.15, imgsz=640, max_dets=2):

    image_dir = Path(image_dir)
    output_dir = Path(output_dir)
    labels_dir = output_dir / "labels"
    labels_dir.mkdir(parents=True, exist_ok=True)

    print("加载 YOLO 模型...")
    model = YOLO(model_path)
    print("✅ 模型加载完成")

    image_files = [f for f in sorted(os.listdir(image_dir)) if os.path.splitext(f.lower())[1] in IMG_EXTENSIONS]
    if not image_files:
        print(f"❌ 未找到图像文件：{image_dir}")
        return

    print(f"发现 {len(image_files)} 张图像，开始推理并保存 YOLO-OBB txt (最多保留 {max_dets} 个检测)...")

    for img_filename in image_files:
        img_path = image_dir / img_filename
        stem = Path(img_filename).stem
        txt_out_path = labels_dir / f"{stem}.txt"

        img = cv2.imread(str(img_path))
        if img is None:
            print(f"❌ 无法读取图像，跳过: {img_path}")
            open(txt_out_path, "w").close()
            continue

        H, W = img.shape[:2]

        results = model(img, save=False, imgsz=imgsz, conf=conf_thresh, mode='obb')
        result = results[0]

        boxes = result.obb
        lines_to_write = []

        if boxes is None or len(boxes) == 0:
            open(txt_out_path, "w").close()
            print(f"ℹ 无检测: {img_filename} -> 生成空 txt")
            continue

        # ==================== 新增：按置信度排序，只取 top-k ====================
        detections = []
        for box in boxes:
            try:
                conf = float(box.conf.cpu().numpy())
                detections.append((conf, box))
            except Exception as e:
                print(f"⚠ 跳过无效 box: {e}")
                continue

        # 按置信度降序排序
        detections.sort(key=lambda x: x[0], reverse=True)
        top_boxes = [box for _, box in detections[:max_dets]]  # 最多取 max_dets 个
        # =====================================================================

        for i, box in enumerate(top_boxes):
            try:
                xywhr = box.xywhr.cpu().numpy()
                if xywhr.ndim == 2:
                    xywhr = xywhr[0]
                cx, cy, bw, bh, r_rad = map(float, xywhr)
            except:
                try:
                    xywh = box.xywh.cpu().numpy()
                    if xywh.ndim == 2:
                        xywh = xywh[0]
                    cx, cy, bw, bh = map(float, xywh)
                    r_rad = 0.0
                except:
                    print(f"⚠ 无法解析 box，跳过 {i}")
                    continue

            # 归一化坐标检测（反归一化为像素）
            if cx <= 1 and cy <= 1 and bw <= 1 and bh <= 1:
                cx *= W
                cy *= H
                bw *= W
                bh *= H

            try:
                cls = int(box.cls.cpu().numpy()[0])
            except:
                cls = 0

            pts = rotate_box_to_corners(cx, cy, bw, bh, r_rad)
            pts = reorder_points_fixed(pts)

            pts_norm = []
            for (x, y) in pts:
                xn = min(max(x / W, 0), 1)
                yn = min(max(y / H, 0), 1)
                pts_norm.extend([xn, yn])

            line = str(cls) + " " + " ".join(f"{v:.6f}" for v in pts_norm)
            lines_to_write.append(line)

        with open(txt_out_path, "w") as f:
            if lines_to_write:
                f.write("\n".join(lines_to_write) + "\n")

        print(f"✅ {img_filename} -> {txt_out_path}  (检测 {len(lines_to_write)} 个)")

    print("\n全部处理完成，结果保存在：", labels_dir)


if __name__ == "__main__":
    MODEL_PATH = r'/home/hx/yolo/ultralytics_yolo11-main/runs/train/exp_obb_new2/weights/best.pt'
    IMAGE_SOURCE_DIR = r"/media/hx/04e879fa-d697-4b02-ac7e-a4148876ebb0/dataset/1/zjdata17"
    OUTPUT_DIR = "/media/hx/04e879fa-d697-4b02-ac7e-a4148876ebb0/dataset/1/labels"

    process_obb_and_save_yolo_txt(
        model_path=MODEL_PATH,
        image_dir=IMAGE_SOURCE_DIR,
        output_dir=OUTPUT_DIR,
        conf_thresh=0.15,
        imgsz=640,
        max_dets=2  # 👈 关键参数：最多保留 2 个检测
    )