Reference for ultralytics/models/rtdetr/predict.py#

class RTDETRPredictor(BasePredictor):
    """RT-DETR (Real-Time Detection Transformer) Predictor extending the BasePredictor class for making predictions.

    This class leverages Vision Transformers to provide real-time object detection while maintaining high accuracy. It
    supports key features like efficient hybrid encoding and IoU-aware query selection.

    Attributes:
        imgsz (int): Image size for inference (must be square and scale-filled).
        args (dict): Argument overrides for the predictor.
        model (torch.nn.Module): The loaded RT-DETR model.
        batch (list): Current batch of processed inputs.

    Methods:
        postprocess: Postprocess raw model predictions to generate bounding boxes and confidence scores.
        pre_transform: Pre-transform input images before feeding them into the model for inference.

    Examples:
        >>> from ultralytics.utils import ASSETS
        >>> from ultralytics.models.rtdetr import RTDETRPredictor
        >>> args = dict(model="rtdetr-l.pt", source=ASSETS)
        >>> predictor = RTDETRPredictor(overrides=args)
        >>> predictor.predict_cli()
    """

def postprocess(self, preds, img, orig_imgs):
    """Postprocess the raw predictions from the model to generate bounding boxes and confidence scores.

    The method filters detections based on confidence and class if specified in `self.args`. It converts model
    predictions (already top-k selected by the decoder head) to Results objects containing properly scaled bounding
    boxes.

    Args:
        preds (list | tuple): List of [predictions, extra] from the model, where predictions have shape (bs,
            num_queries, 6) with format [cx, cy, w, h, score, class].
        img (torch.Tensor): Processed input images with shape (N, 3, H, W).
        orig_imgs (list | torch.Tensor): Original, unprocessed images.

    Returns:
        (list[Results]): A list of Results objects containing the post-processed bounding boxes, confidence scores,
            and class labels.
    """
    if isinstance(preds, (list, tuple)):
        preds = preds[0]
    bboxes, scores, labels = preds.split((4, 1, 1), dim=-1)
    if not isinstance(orig_imgs, list):  # input images are a torch.Tensor, not a list
        orig_imgs = ops.convert_torch2numpy_batch(orig_imgs)[..., ::-1]

    results = []
    for bbox, score, label, orig_img, img_path in zip(bboxes, scores, labels, orig_imgs, self.batch[0]):
        bbox = ops.xywh2xyxy(bbox)
        idx = score.squeeze(-1) > self.args.conf
        if self.args.classes is not None:
            idx = (label == torch.tensor(self.args.classes, device=label.device)).any(1) & idx
        pred = torch.cat([bbox, score, label], dim=-1)[idx][: self.args.max_det]
        oh, ow = orig_img.shape[:2]
        pred[..., [0, 2]] *= ow  # scale x coordinates to original width
        pred[..., [1, 3]] *= oh  # scale y coordinates to original height
        results.append(Results(orig_img, path=img_path, names=self.model.names, boxes=pred))
    return results

def pre_transform(self, im):
    """Pre-transform input images before feeding them into the model for inference.

    The input images are letterboxed to ensure a square aspect ratio and scale-filled.

    Args:
        im (list[np.ndarray]): Input images of shape [(H, W, 3) x N].

    Returns:
        (list): List of pre-transformed images ready for model inference.
    """
    letterbox = LetterBox(self.imgsz, auto=False, scale_fill=True)
    return [letterbox(image=x) for x in im]