由百度开发的实时检测 Transformer (RT-DETR) 是一种先进的端到端对象检测器，可在保持高精度的同时提供实时性能。它基于 DETR（无 NMS 框架）的思想，同时引入了基于 conv 的骨干网络和一个高效的混合编码器以获得实时速度。

RT-DETR 通过解耦尺度内交互和跨尺度融合来高效地处理多尺度特征。该模型具有高度的适应性，支持使用不同的解码器层灵活调整推理速度，而无需重新训练。RT-DETR 在 CUDA 与 TensorRT 等加速后端上表现出色，优于许多其他实时对象检测器。

Ultralytics Python API 提供了具有不同规模的预训练 PaddlePaddle RT-DETR 模型，本文使用Ultralytics进行RT-DETR的复现。

训练没有任何问题。但是评估的时候，如果开启save_json=True的选项，会直接报错：

(Pytorch) PS D:\CodeProject\datasets\VisDrone> python .\ultralytics-main\check.py
Ultralytics 8.3.192  Python-3.10.18 torch-2.0.0 CUDA:0 (NVIDIA GeForce RTX 4090, 24564MiB)
rtdetr-l summary: 302 layers, 32,004,290 parameters, 0 gradients, 103.5 GFLOPs
val: Fast image access  (ping: 0.00.0 ms, read: 2287.3481.3 MB/s, size: 175.5 KB)
val: Scanning D:\CodeProject\datasets\VisDrone\VisDrone_YOLO\VisDrone2019-DET-val\labels.cache... 548 images, 0 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 548/548  0.0s
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 0% ──────────── 0/137  0.1s<
Traceback (most recent call last):
  File "D:\CodeProject\datasets\VisDrone\ultralytics-main\check.py", line 11, in <module>
    result = model.val(data=r'D:\CodeProject\datasets\VisDrone\ultralytics-main\ultralytics\cfg\datasets\VisDrone.yaml',
  File "D:\CodeProject\datasets\VisDrone\ultralytics-main\ultralytics\engine\model.py", line 635, in val
    validator(model=self.model)
  File "D:\Anaconda3\envs\Pytorch\lib\site-packages\torch\utils\_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "D:\CodeProject\datasets\VisDrone\ultralytics-main\ultralytics\engine\validator.py", line 221, in __call__
    self.update_metrics(preds, batch)
  File "D:\CodeProject\datasets\VisDrone\ultralytics-main\ultralytics\models\yolo\detect\val.py", line 210, in update_metrics
    predn_scaled = self.scale_preds(predn, pbatch)
  File "D:\CodeProject\datasets\VisDrone\ultralytics-main\ultralytics\models\yolo\detect\val.py", line 407, in scale_preds
    "bboxes": ops.scale_boxes(
  File "D:\CodeProject\datasets\VisDrone\ultralytics-main\ultralytics\utils\ops.py", line 128, in scale_boxes
    gain = ratio_pad[0][0]
TypeError: 'float' object is not subscriptable

这个错误是因为在 ops.scale_boxes() 函数中，ratio_pad 参数被当作元组处理，但实际上传入的是一个浮点数。

需要进行修改：

def scale_boxes(img1_shape, boxes, img0_shape, ratio_pad=None, padding: bool = True, xywh: bool = False):
    """
    Rescale bounding boxes from one image shape to another.

    Rescales bounding boxes from img1_shape to img0_shape, accounting for padding and aspect ratio changes.
    Supports both xyxy and xywh box formats.

    Args:
        img1_shape (tuple): Shape of the source image (height, width).
        boxes (torch.Tensor): Bounding boxes to rescale in format (N, 4).
        img0_shape (tuple): Shape of the target image (height, width).
        ratio_pad (tuple, optional): Tuple of (ratio, pad) for scaling. If None, calculated from image shapes.
        padding (bool): Whether boxes are based on YOLO-style augmented images with padding.
        xywh (bool): Whether box format is xywh (True) or xyxy (False).

    Returns:
        (torch.Tensor): Rescaled bounding boxes in the same format as input.
    """
    if ratio_pad is None:  # calculate from img0_shape
        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
        pad_x = round((img1_shape[1] - img0_shape[1] * gain) / 2 - 0.1)
        pad_y = round((img1_shape[0] - img0_shape[0] * gain) / 2 - 0.1)
    else:
        if isinstance(ratio_pad, (int, float)):
            gain = ratio_pad
            pad_x, pad_y = 0, 0
        elif isinstance(ratio_pad, (list, tuple)):
            if len(ratio_pad) == 2:
                if isinstance(ratio_pad[0], (list, tuple)):
                    gain = ratio_pad[0][0]
                else:
                    gain = ratio_pad[0]

                if isinstance(ratio_pad[1], (list, tuple)) and len(ratio_pad[1]) >= 2:
                    pad_x, pad_y = ratio_pad[1][0], ratio_pad[1][1]
                else:
                    pad_x, pad_y = 0, 0
            else:
                gain = ratio_pad[0] if len(ratio_pad) > 0 else 1.0
                pad_x, pad_y = 0, 0
        else:
            gain = 1.0
            pad_x, pad_y = 0, 0

    if padding:
        boxes[..., 0] -= pad_x  # x padding
        boxes[..., 1] -= pad_y  # y padding
        if not xywh:
            boxes[..., 2] -= pad_x  # x padding
            boxes[..., 3] -= pad_y  # y padding
    boxes[..., :4] /= gain
    return clip_boxes(boxes, img0_shape)

这样可以save_json成功，但是进行coco指标评估的时候，虽然没有报错，但是指标数值很低很低！

python D:\CodeProject\datasets\VisDrone\ultralytics-main\COCO_Evalution.py --annotations D:\CodeProject\datasets\VisDrone\VisDrone_YOLO\instances_val_2017.json --predictions D:\CodeProject\datasets\VisDrone\runs\detect\val9\predictions.json

评估的结果格式会出现问题，所以为了正常使用，我们需要进行逐个评估，并根据GT的格式来生成json文件。代码如下：

from ultralytics import RTDETR
import json
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
import os


def correct_coco_evaluation(model_path, image_dir, gt_json_path, output_dir='results'):
    os.makedirs(output_dir, exist_ok=True)

    coco_gt = COCO(gt_json_path)
    gt_images = coco_gt.imgs
    print(f"GT 中包含 {len(gt_images)} 张图像")

    image_name_to_id = {}
    for img_id, img_info in gt_images.items():
        image_name_to_id[img_info['file_name']] = img_id

    model = RTDETR(model_path)
    results = model.predict(source=image_dir, save=False, conf=0.001, workers=0)

    predictions = []
    pred_id = 1

    for result in results:
        image_path = result.path
        image_name = os.path.basename(image_path)

        if image_name in image_name_to_id:
            image_id = image_name_to_id[image_name]
        else:
            print(f"警告: 未找到图像 {image_name} 在 GT 中的对应ID，跳过")
            continue

        if result.boxes is not None:
            for box in result.boxes:
                xyxy = box.xyxy[0].tolist()
                x1, y1, x2, y2 = xyxy
                width = x2 - x1
                height = y2 - y1

                predictions.append({
                    'id': pred_id,
                    'image_id': image_id,
                    'category_id': int(box.cls.item()) + 1,
                    'bbox': [x1, y1, width, height],
                    'score': box.conf.item(),
                    'area': width * height
                })
                pred_id += 1
        else:
            print(f"图像 {image_name} 未检测到任何目标")

    print(f"生成 {len(predictions)} 个预测")

    pred_file = os.path.join(output_dir, 'predictions.json')
    with open(pred_file, 'w') as f:
        json.dump(predictions, f, indent=2)

    coco_dt = coco_gt.loadRes(pred_file)

    coco_eval = COCOeval(coco_gt, coco_dt, 'bbox')
    coco_eval.evaluate()
    coco_eval.accumulate()
    coco_eval.summarize()

    return coco_eval.stats


if __name__ == '__main__':
    stats = correct_coco_evaluation(
        model_path=r'D:\CodeProject\datasets\VisDrone\ultralytics-main\runs\detect\train4\weights\best.pt',
        image_dir=r'D:\CodeProject\datasets\VisDrone\VisDrone_YOLO\VisDrone2019-DET-test\images',
        gt_json_path=r'D:\CodeProject\datasets\VisDrone\ultralytics-main\instances_test_2017.json'
    )

    print(f"\nCOCO 评估结果:")
    print(f"AP@[0.5:0.95]: {stats[0]:.3f}")
    print(f"AP@0.5: {stats[1]:.3f}")
    print(f"AP@0.75: {stats[2]:.3f}")
    print(f"AP_small: {stats[3]:.3f}")
    print(f"AP_medium: {stats[4]:.3f}")
    print(f"AP_large: {stats[5]:.3f}")

结果顺利：