ultralytics-opencv-headless 8.3.242__py3-none-any.whl

ultralytics/models/fastsam/predict.py

@@ -0,0 +1,169 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+import torch
+from PIL import Image
+
+from ultralytics.models.yolo.segment import SegmentationPredictor
+from ultralytics.utils import DEFAULT_CFG
+from ultralytics.utils.metrics import box_iou
+from ultralytics.utils.ops import scale_masks
+from ultralytics.utils.torch_utils import TORCH_1_10
+
+from .utils import adjust_bboxes_to_image_border
+
+
+class FastSAMPredictor(SegmentationPredictor):
+    """FastSAMPredictor is specialized for fast SAM (Segment Anything Model) segmentation prediction tasks.
+
+    This class extends the SegmentationPredictor, customizing the prediction pipeline specifically for fast SAM. It
+    adjusts post-processing steps to incorporate mask prediction and non-maximum suppression while optimizing for
+    single-class segmentation.
+
+    Attributes:
+        prompts (dict): Dictionary containing prompt information for segmentation (bboxes, points, labels, texts).
+        device (torch.device): Device on which model and tensors are processed.
+        clip (Any, optional): CLIP model used for text-based prompting, loaded on demand.
+
+    Methods:
+        postprocess: Apply postprocessing to FastSAM predictions and handle prompts.
+        prompt: Perform image segmentation inference based on various prompt types.
+        set_prompts: Set prompts to be used during inference.
+    """
+
+    def __init__(self, cfg=DEFAULT_CFG, overrides=None, _callbacks=None):
+        """Initialize the FastSAMPredictor with configuration and callbacks.
+
+        This initializes a predictor specialized for Fast SAM (Segment Anything Model) segmentation tasks. The predictor
+        extends SegmentationPredictor with custom post-processing for mask prediction and non-maximum suppression
+        optimized for single-class segmentation.
+
+        Args:
+            cfg (dict): Configuration for the predictor.
+            overrides (dict, optional): Configuration overrides.
+            _callbacks (list, optional): List of callback functions.
+        """
+        super().__init__(cfg, overrides, _callbacks)
+        self.prompts = {}
+
+    def postprocess(self, preds, img, orig_imgs):
+        """Apply postprocessing to FastSAM predictions and handle prompts.
+
+        Args:
+            preds (list[torch.Tensor]): Raw predictions from the model.
+            img (torch.Tensor): Input image tensor that was fed to the model.
+            orig_imgs (list[np.ndarray]): Original images before preprocessing.
+
+        Returns:
+            (list[Results]): Processed results with prompts applied.
+        """
+        bboxes = self.prompts.pop("bboxes", None)
+        points = self.prompts.pop("points", None)
+        labels = self.prompts.pop("labels", None)
+        texts = self.prompts.pop("texts", None)
+        results = super().postprocess(preds, img, orig_imgs)
+        for result in results:
+            full_box = torch.tensor(
+                [0, 0, result.orig_shape[1], result.orig_shape[0]], device=preds[0].device, dtype=torch.float32
+            )
+            boxes = adjust_bboxes_to_image_border(result.boxes.xyxy, result.orig_shape)
+            idx = torch.nonzero(box_iou(full_box[None], boxes) > 0.9).flatten()
+            if idx.numel() != 0:
+                result.boxes.xyxy[idx] = full_box
+
+        return self.prompt(results, bboxes=bboxes, points=points, labels=labels, texts=texts)
+
+    def prompt(self, results, bboxes=None, points=None, labels=None, texts=None):
+        """Perform image segmentation inference based on cues like bounding boxes, points, and text prompts.
+
+        Args:
+            results (Results | list[Results]): Original inference results from FastSAM models without any prompts.
+            bboxes (np.ndarray | list, optional): Bounding boxes with shape (N, 4), in XYXY format.
+            points (np.ndarray | list, optional): Points indicating object locations with shape (N, 2), in pixels.
+            labels (np.ndarray | list, optional): Labels for point prompts, shape (N, ). 1 = foreground, 0 = background.
+            texts (str | list[str], optional): Textual prompts, a list containing string objects.
+
+        Returns:
+            (list[Results]): Output results filtered and determined by the provided prompts.
+        """
+        if bboxes is None and points is None and texts is None:
+            return results
+        prompt_results = []
+        if not isinstance(results, list):
+            results = [results]
+        for result in results:
+            if len(result) == 0:
+                prompt_results.append(result)
+                continue
+            masks = result.masks.data
+            if masks.shape[1:] != result.orig_shape:
+                masks = (scale_masks(masks[None].float(), result.orig_shape)[0] > 0.5).byte()
+            # bboxes prompt
+            idx = torch.zeros(len(result), dtype=torch.bool, device=self.device)
+            if bboxes is not None:
+                bboxes = torch.as_tensor(bboxes, dtype=torch.int32, device=self.device)
+                bboxes = bboxes[None] if bboxes.ndim == 1 else bboxes
+                bbox_areas = (bboxes[:, 3] - bboxes[:, 1]) * (bboxes[:, 2] - bboxes[:, 0])
+                mask_areas = torch.stack([masks[:, b[1] : b[3], b[0] : b[2]].sum(dim=(1, 2)) for b in bboxes])
+                full_mask_areas = torch.sum(masks, dim=(1, 2))
+
+                union = bbox_areas[:, None] + full_mask_areas - mask_areas
+                idx[torch.argmax(mask_areas / union, dim=1)] = True
+            if points is not None:
+                points = torch.as_tensor(points, dtype=torch.int32, device=self.device)
+                points = points[None] if points.ndim == 1 else points
+                if labels is None:
+                    labels = torch.ones(points.shape[0])
+                labels = torch.as_tensor(labels, dtype=torch.int32, device=self.device)
+                assert len(labels) == len(points), (
+                    f"Expected `labels` to have the same length as `points`, but got {len(labels)} and {len(points)}."
+                )
+                point_idx = (
+                    torch.ones(len(result), dtype=torch.bool, device=self.device)
+                    if labels.sum() == 0  # all negative points
+                    else torch.zeros(len(result), dtype=torch.bool, device=self.device)
+                )
+                for point, label in zip(points, labels):
+                    point_idx[torch.nonzero(masks[:, point[1], point[0]], as_tuple=True)[0]] = bool(label)
+                idx |= point_idx
+            if texts is not None:
+                if isinstance(texts, str):
+                    texts = [texts]
+                crop_ims, filter_idx = [], []
+                for i, b in enumerate(result.boxes.xyxy.tolist()):
+                    x1, y1, x2, y2 = (int(x) for x in b)
+                    if (masks[i].sum() if TORCH_1_10 else masks[i].sum(0).sum()) <= 100:  # torch 1.9 bug workaround
+                        filter_idx.append(i)
+                        continue
+                    crop_ims.append(Image.fromarray(result.orig_img[y1:y2, x1:x2, ::-1]))
+                similarity = self._clip_inference(crop_ims, texts)
+                text_idx = torch.argmax(similarity, dim=-1)  # (M, )
+                if len(filter_idx):
+                    text_idx += (torch.tensor(filter_idx, device=self.device)[None] <= int(text_idx)).sum(0)
+                idx[text_idx] = True
+
+            prompt_results.append(result[idx])
+
+        return prompt_results
+
+    def _clip_inference(self, images, texts):
+        """Perform CLIP inference to calculate similarity between images and text prompts.
+
+        Args:
+            images (list[PIL.Image]): List of source images, each should be PIL.Image with RGB channel order.
+            texts (list[str]): List of prompt texts, each should be a string object.
+
+        Returns:
+            (torch.Tensor): Similarity matrix between given images and texts with shape (M, N).
+        """
+        from ultralytics.nn.text_model import CLIP
+
+        if not hasattr(self, "clip"):
+            self.clip = CLIP("ViT-B/32", device=self.device)
+        images = torch.stack([self.clip.image_preprocess(image).to(self.device) for image in images])
+        image_features = self.clip.encode_image(images)
+        text_features = self.clip.encode_text(self.clip.tokenize(texts))
+        return text_features @ image_features.T  # (M, N)
+
+    def set_prompts(self, prompts):
+        """Set prompts to be used during inference."""
+        self.prompts = prompts

ultralytics/models/fastsam/utils.py

@@ -0,0 +1,23 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+
+def adjust_bboxes_to_image_border(boxes, image_shape, threshold=20):
+    """Adjust bounding boxes to stick to image border if they are within a certain threshold.
+
+    Args:
+        boxes (torch.Tensor): Bounding boxes with shape (N, 4) in xyxy format.
+        image_shape (tuple): Image dimensions as (height, width).
+        threshold (int): Pixel threshold for considering a box close to the border.
+
+    Returns:
+        (torch.Tensor): Adjusted bounding boxes with shape (N, 4).
+    """
+    # Image dimensions
+    h, w = image_shape
+
+    # Adjust boxes that are close to image borders
+    boxes[boxes[:, 0] < threshold, 0] = 0  # x1
+    boxes[boxes[:, 1] < threshold, 1] = 0  # y1
+    boxes[boxes[:, 2] > w - threshold, 2] = w  # x2
+    boxes[boxes[:, 3] > h - threshold, 3] = h  # y2
+    return boxes

ultralytics/models/fastsam/val.py

@@ -0,0 +1,38 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+from ultralytics.models.yolo.segment import SegmentationValidator
+
+
+class FastSAMValidator(SegmentationValidator):
+    """Custom validation class for FastSAM (Segment Anything Model) segmentation in the Ultralytics YOLO framework.
+
+    Extends the SegmentationValidator class, customizing the validation process specifically for FastSAM. This class
+    sets the task to 'segment' and uses the SegmentMetrics for evaluation. Additionally, plotting features are disabled
+    to avoid errors during validation.
+
+    Attributes:
+        dataloader (torch.utils.data.DataLoader): The data loader object used for validation.
+        save_dir (Path): The directory where validation results will be saved.
+        args (SimpleNamespace): Additional arguments for customization of the validation process.
+        _callbacks (list): List of callback functions to be invoked during validation.
+        metrics (SegmentMetrics): Segmentation metrics calculator for evaluation.
+
+    Methods:
+        __init__: Initialize the FastSAMValidator with custom settings for FastSAM.
+    """
+
+    def __init__(self, dataloader=None, save_dir=None, args=None, _callbacks=None):
+        """Initialize the FastSAMValidator class, setting the task to 'segment' and metrics to SegmentMetrics.
+
+        Args:
+            dataloader (torch.utils.data.DataLoader, optional): DataLoader to be used for validation.
+            save_dir (Path, optional): Directory to save results.
+            args (SimpleNamespace, optional): Configuration for the validator.
+            _callbacks (list, optional): List of callback functions to be invoked during validation.
+
+        Notes:
+            Plots for ConfusionMatrix and other related metrics are disabled in this class to avoid errors.
+        """
+        super().__init__(dataloader, save_dir, args, _callbacks)
+        self.args.task = "segment"
+        self.args.plots = False  # disable ConfusionMatrix and other plots to avoid errors

ultralytics/models/nas/__init__.py

@@ -0,0 +1,7 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+from .model import NAS
+from .predict import NASPredictor
+from .val import NASValidator
+
+__all__ = "NAS", "NASPredictor", "NASValidator"

ultralytics/models/nas/model.py

@@ -0,0 +1,98 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import torch
+
+from ultralytics.engine.model import Model
+from ultralytics.utils import DEFAULT_CFG_DICT
+from ultralytics.utils.downloads import attempt_download_asset
+from ultralytics.utils.patches import torch_load
+from ultralytics.utils.torch_utils import model_info
+
+from .predict import NASPredictor
+from .val import NASValidator
+
+
+class NAS(Model):
+    """YOLO-NAS model for object detection.
+
+    This class provides an interface for the YOLO-NAS models and extends the `Model` class from Ultralytics engine. It
+    is designed to facilitate the task of object detection using pre-trained or custom-trained YOLO-NAS models.
+
+    Attributes:
+        model (torch.nn.Module): The loaded YOLO-NAS model.
+        task (str): The task type for the model, defaults to 'detect'.
+        predictor (NASPredictor): The predictor instance for making predictions.
+        validator (NASValidator): The validator instance for model validation.
+
+    Methods:
+        info: Log model information and return model details.
+
+    Examples:
+        >>> from ultralytics import NAS
+        >>> model = NAS("yolo_nas_s")
+        >>> results = model.predict("ultralytics/assets/bus.jpg")
+
+    Notes:
+        YOLO-NAS models only support pre-trained models. Do not provide YAML configuration files.
+    """
+
+    def __init__(self, model: str = "yolo_nas_s.pt") -> None:
+        """Initialize the NAS model with the provided or default model."""
+        assert Path(model).suffix not in {".yaml", ".yml"}, "YOLO-NAS models only support pre-trained models."
+        super().__init__(model, task="detect")
+
+    def _load(self, weights: str, task=None) -> None:
+        """Load an existing NAS model weights or create a new NAS model with pretrained weights.
+
+        Args:
+            weights (str): Path to the model weights file or model name.
+            task (str, optional): Task type for the model.
+        """
+        import super_gradients
+
+        suffix = Path(weights).suffix
+        if suffix == ".pt":
+            self.model = torch_load(attempt_download_asset(weights))
+        elif suffix == "":
+            self.model = super_gradients.training.models.get(weights, pretrained_weights="coco")
+
+        # Override the forward method to ignore additional arguments
+        def new_forward(x, *args, **kwargs):
+            """Ignore additional __call__ arguments."""
+            return self.model._original_forward(x)
+
+        self.model._original_forward = self.model.forward
+        self.model.forward = new_forward
+
+        # Standardize model attributes for compatibility
+        self.model.fuse = lambda verbose=True: self.model
+        self.model.stride = torch.tensor([32])
+        self.model.names = dict(enumerate(self.model._class_names))
+        self.model.is_fused = lambda: False  # for info()
+        self.model.yaml = {}  # for info()
+        self.model.pt_path = weights  # for export()
+        self.model.task = "detect"  # for export()
+        self.model.args = {**DEFAULT_CFG_DICT, **self.overrides}  # for export()
+        self.model.eval()
+
+    def info(self, detailed: bool = False, verbose: bool = True) -> dict[str, Any]:
+        """Log model information.
+
+        Args:
+            detailed (bool): Show detailed information about model.
+            verbose (bool): Controls verbosity.
+
+        Returns:
+            (dict[str, Any]): Model information dictionary.
+        """
+        return model_info(self.model, detailed=detailed, verbose=verbose, imgsz=640)
+
+    @property
+    def task_map(self) -> dict[str, dict[str, Any]]:
+        """Return a dictionary mapping tasks to respective predictor and validator classes."""
+        return {"detect": {"predictor": NASPredictor, "validator": NASValidator}}

ultralytics/models/nas/predict.py

@@ -0,0 +1,56 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+import torch
+
+from ultralytics.models.yolo.detect.predict import DetectionPredictor
+from ultralytics.utils import ops
+
+
+class NASPredictor(DetectionPredictor):
+    """Ultralytics YOLO NAS Predictor for object detection.
+
+    This class extends the DetectionPredictor from Ultralytics engine and is responsible for post-processing the raw
+    predictions generated by the YOLO NAS models. It applies operations like non-maximum suppression and scaling the
+    bounding boxes to fit the original image dimensions.
+
+    Attributes:
+        args (Namespace): Namespace containing various configurations for post-processing including confidence
+            threshold, IoU threshold, agnostic NMS flag, maximum detections, and class filtering options.
+        model (torch.nn.Module): The YOLO NAS model used for inference.
+        batch (list): Batch of inputs for processing.
+
+    Examples:
+        >>> from ultralytics import NAS
+        >>> model = NAS("yolo_nas_s")
+        >>> predictor = model.predictor
+
+        Assume that raw_preds, img, orig_imgs are available
+        >>> results = predictor.postprocess(raw_preds, img, orig_imgs)
+
+    Notes:
+        Typically, this class is not instantiated directly. It is used internally within the NAS class.
+    """
+
+    def postprocess(self, preds_in, img, orig_imgs):
+        """Postprocess NAS model predictions to generate final detection results.
+
+        This method takes raw predictions from a YOLO NAS model, converts bounding box formats, and applies
+        post-processing operations to generate the final detection results compatible with Ultralytics result
+        visualization and analysis tools.
+
+        Args:
+            preds_in (list): Raw predictions from the NAS model, typically containing bounding boxes and class scores.
+            img (torch.Tensor): Input image tensor that was fed to the model, with shape (B, C, H, W).
+            orig_imgs (list | torch.Tensor | np.ndarray): Original images before preprocessing, used for scaling
+                coordinates back to original dimensions.
+
+        Returns:
+            (list): List of Results objects containing the processed predictions for each image in the batch.
+
+        Examples:
+            >>> predictor = NAS("yolo_nas_s").predictor
+            >>> results = predictor.postprocess(raw_preds, img, orig_imgs)
+        """
+        boxes = ops.xyxy2xywh(preds_in[0][0])  # Convert bounding boxes from xyxy to xywh format
+        preds = torch.cat((boxes, preds_in[0][1]), -1).permute(0, 2, 1)  # Concatenate boxes with class scores
+        return super().postprocess(preds, img, orig_imgs)

ultralytics/models/nas/val.py

@@ -0,0 +1,38 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+import torch
+
+from ultralytics.models.yolo.detect import DetectionValidator
+from ultralytics.utils import ops
+
+__all__ = ["NASValidator"]
+
+
+class NASValidator(DetectionValidator):
+    """Ultralytics YOLO NAS Validator for object detection.
+
+    Extends DetectionValidator from the Ultralytics models package and is designed to post-process the raw predictions
+    generated by YOLO NAS models. It performs non-maximum suppression to remove overlapping and low-confidence boxes,
+    ultimately producing the final detections.
+
+    Attributes:
+        args (Namespace): Namespace containing various configurations for post-processing, such as confidence and IoU
+            thresholds.
+        lb (torch.Tensor): Optional tensor for multilabel NMS.
+
+    Examples:
+        >>> from ultralytics import NAS
+        >>> model = NAS("yolo_nas_s")
+        >>> validator = model.validator
+        >>> # Assumes that raw_preds are available
+        >>> final_preds = validator.postprocess(raw_preds)
+
+    Notes:
+        This class is generally not instantiated directly but is used internally within the NAS class.
+    """
+
+    def postprocess(self, preds_in):
+        """Apply Non-maximum suppression to prediction outputs."""
+        boxes = ops.xyxy2xywh(preds_in[0][0])  # Convert bounding box format from xyxy to xywh
+        preds = torch.cat((boxes, preds_in[0][1]), -1).permute(0, 2, 1)  # Concatenate boxes with scores and permute
+        return super().postprocess(preds)

ultralytics/models/rtdetr/__init__.py

@@ -0,0 +1,7 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+from .model import RTDETR
+from .predict import RTDETRPredictor
+from .val import RTDETRValidator
+
+__all__ = "RTDETR", "RTDETRPredictor", "RTDETRValidator"

ultralytics/models/rtdetr/model.py

@@ -0,0 +1,63 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+"""
+Interface for Baidu's RT-DETR, a Vision Transformer-based real-time object detector.
+
+RT-DETR offers real-time performance and high accuracy, excelling in accelerated backends like CUDA with TensorRT.
+It features an efficient hybrid encoder and IoU-aware query selection for enhanced detection accuracy.
+
+References:
+    https://arxiv.org/pdf/2304.08069.pdf
+"""
+
+from ultralytics.engine.model import Model
+from ultralytics.nn.tasks import RTDETRDetectionModel
+from ultralytics.utils.torch_utils import TORCH_1_11
+
+from .predict import RTDETRPredictor
+from .train import RTDETRTrainer
+from .val import RTDETRValidator
+
+
+class RTDETR(Model):
+    """Interface for Baidu's RT-DETR model, a Vision Transformer-based real-time object detector.
+
+    This model provides real-time performance with high accuracy. It supports efficient hybrid encoding, IoU-aware query
+    selection, and adaptable inference speed.
+
+    Attributes:
+        model (str): Path to the pre-trained model.
+
+    Methods:
+        task_map: Return a task map for RT-DETR, associating tasks with corresponding Ultralytics classes.
+
+    Examples:
+        Initialize RT-DETR with a pre-trained model
+        >>> from ultralytics import RTDETR
+        >>> model = RTDETR("rtdetr-l.pt")
+        >>> results = model("image.jpg")
+    """
+
+    def __init__(self, model: str = "rtdetr-l.pt") -> None:
+        """Initialize the RT-DETR model with the given pre-trained model file.
+
+        Args:
+            model (str): Path to the pre-trained model. Supports .pt, .yaml, and .yml formats.
+        """
+        assert TORCH_1_11, "RTDETR requires torch>=1.11"
+        super().__init__(model=model, task="detect")
+
+    @property
+    def task_map(self) -> dict:
+        """Return a task map for RT-DETR, associating tasks with corresponding Ultralytics classes.
+
+        Returns:
+            (dict): A dictionary mapping task names to Ultralytics task classes for the RT-DETR model.
+        """
+        return {
+            "detect": {
+                "predictor": RTDETRPredictor,
+                "validator": RTDETRValidator,
+                "trainer": RTDETRTrainer,
+                "model": RTDETRDetectionModel,
+            }
+        }

ultralytics/models/rtdetr/predict.py

@@ -0,0 +1,88 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+import torch
+
+from ultralytics.data.augment import LetterBox
+from ultralytics.engine.predictor import BasePredictor
+from ultralytics.engine.results import Results
+from ultralytics.utils import ops
+
+
+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 to Results objects containing properly scaled bounding boxes.
+
+        Args:
+            preds (list | tuple): List of [predictions, extra] from the model, where predictions contain bounding boxes
+                and scores.
+            img (torch.Tensor): Processed input images with shape (N, 3, H, W).
+            orig_imgs (list | torch.Tensor): Original, unprocessed images.
+
+        Returns:
+            results (list[Results]): A list of Results objects containing the post-processed bounding boxes, confidence
+                scores, and class labels.
+        """
+        if not isinstance(preds, (list, tuple)):  # list for PyTorch inference but list[0] Tensor for export inference
+            preds = [preds, None]
+
+        nd = preds[0].shape[-1]
+        bboxes, scores = preds[0].split((4, nd - 4), 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, orig_img, img_path in zip(bboxes, scores, orig_imgs, self.batch[0]):  # (300, 4)
+            bbox = ops.xywh2xyxy(bbox)
+            max_score, cls = score.max(-1, keepdim=True)  # (300, 1)
+            idx = max_score.squeeze(-1) > self.args.conf  # (300, )
+            if self.args.classes is not None:
+                idx = (cls == torch.tensor(self.args.classes, device=cls.device)).any(1) & idx
+            pred = torch.cat([bbox, max_score, cls], dim=-1)[idx]  # filter
+            pred = pred[pred[:, 4].argsort(descending=True)][: 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] | torch.Tensor): Input images of shape (N, 3, H, W) for tensor, [(H, W, 3) x N] for
+                list.
+
+        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]