* @brief Softmax function for DFL */

Run the complete OBB inference pipeline on a single image. This method orchestrates the full workflow: - Pre-process the inp

Run BPU inference for one input image.

Run the full classification pipeline on one image.

Set BPU scheduling parameters like priority and core affinity. Args: priority (Optional[int]): Scheduling priority in ra

* @brief Draw detection results on image */

Resolve a path to an absolute path. Args: path: The path string to resolve. base_dir: Optional base directory for relative resolu

解析路径，支持相对路径和绝对路径

Save an image to the target path. Args: path (str): Output image path. image (np.ndarray): Image array to save. Raises:

Set optional BPU scheduling parameters.

Resize a grayscale image using letterbox (aspect-ratio preserving) strategy. This function resizes a grayscale image while preserving its aspect

Run the complete HGNetV2 inference pipeline on one image. Args: image: Input image in BGR format. resize_typ

Save an image to the target path. Args: path: Output image path. image: Image array to save. Raises: RuntimeError: I

Write a minimal index.rst with a toctree.

Returns list of utf-8 byte and a corresponding list of unicode strings. The reversible bpe codes work on unicode strings. This means you

Decode a sequence of CTC indices into a string. Args: t (np.ndarray): Encoded indices. length (np.ndarray): Length of

Draw bounding boxes with class names and scores.

Execute inference on BPU using hbm_runtime. Args: input_tensor (Dict[str, Dict[str, np.ndarray]]): Prepared input tensor

Execute inference on BPU using hbm_runtime. Args: input_tensor (Dict[str, Dict[str, np.ndarray]]): Prepared input tensor

Execute inference on BPU using hbm_runtime. Args: input_tensor (Dict[str, Dict[str, np.ndarray]]): Prepared input tensor

Execute inference on BPU using hbm_runtime. Args: input_tensor (Dict[str, Dict[str, np.ndarray]]): Prepared input tensor

Execute inference on BPU using hbm_runtime. Args: input_tensor (Dict[str, Dict[str, np.ndarray]]): Prepared input tensor

Generate anchor center positions on a square grid. This function generates evenly spaced anchor center coordinates for a square grid of size

Return set of symbol pairs in a word. Word is represented as tuple of symbols (symbols being variable-length strings).

Replace a Linear layer with an equivalent 1x1 Conv2d for BPU.

Convert raw output tensor to Top-K classification results. Args: outputs (Dict[str, Dict[str, np.ndarray]]): Raw model o

Convert raw outputs to final pose estimation results. Args: outputs (Dict[str, Dict[str, np.ndarray]]): Raw model output

Convert raw model outputs to final detection results. Args: outputs (Dict[str, Dict[str, np.ndarray]]): Raw model output

Convert raw model outputs to final segmentation results. Args: outputs (Dict[str, Dict[str, np.ndarray]]): Raw model out

Convert raw outputs to final rotated box results. Args: outputs (Dict[str, Dict[str, np.ndarray]]): Raw model outputs.

Convert a BGR image to packed NV12 input for hbm_runtime. Args: img (np.ndarray): Input image in BGR format.

Convert a BGR image to packed NV12 input for hbm_runtime. Args: img (np.ndarray): Input image in BGR format.

Convert a BGR image to packed NV12 input for hbm_runtime. Args: img (np.ndarray): Input image in BGR format.

Convert a BGR image to packed NV12 input for hbm_runtime. Args: img (np.ndarray): Input image in BGR format.

Convert a BGR image to packed NV12 input for hbm_runtime. Args: img (np.ndarray): Input image in BGR format.

Run the complete EfficientViT inference pipeline on one image. Args: image: Input image in BGR format. resiz

Run the complete EdgeNeXt inference pipeline on one image. Args: image: Input image in BGR format. resize_ty

Run the complete LPRNet pipeline on one binary input tensor. Returns: str: Final decoded license plate string produced b

High-level interface for one-click inference. Args: image (np.ndarray): Input image in BGR format. Returns:

Run the complete ResNet inference pipeline on one image. Args: image: Input image in BGR format. resize_type

Run the complete RepVGG inference pipeline on one image. Args: image: Input image in BGR format. resize_type

Run the complete MobileNetV3 inference pipeline on one image. Args: image: Input image in BGR format. resize

Run the complete CLIP image-text matching pipeline. Args: image: Input image in OpenCV BGR format. texts: Ca

Run the complete MobileNetV2 inference pipeline on one image. Args: image: Input image in BGR format. resize

Run the complete VargConvNet inference pipeline on one image. Args: image: Input image in BGR format. resize

Run the complete segmentation pipeline on a single image. Args: img: Input image in BGR format. image_format: Input i

Run the complete OCR pipeline: detect → crop → recognize. Args: image (np.ndarray): Input image in BGR format. Returns:

Run the complete MobileNetV1 inference pipeline on one image. Args: image: Input image in BGR format. resize

Execute the full FCOS pipeline on one image. Args: image (np.ndarray): Input BGR image. Returns: Tu

Run the complete RepGhost inference pipeline on one image. Args: image: Input image in BGR format. resize_ty

Run the complete EfficientFormer inference pipeline on one image. Args: image: Input image in BGR format. re

Run the complete MobileOne inference pipeline on one image. Args: image: Input image in BGR format. resize_t

Run the complete YOLOWorld detection pipeline on one image. Args: image: Input image in OpenCV BGR format. p

Run the complete FasterNet inference pipeline on one image. Args: image: Input image in BGR format. resize_t

Run the complete EfficientNet inference pipeline on one image. Args: image: Input image in BGR format. resiz

Run the complete FastViT inference pipeline on one image. Args: image: Input image in BGR format. resize_typ

Run the complete ResNeXt inference pipeline on one image. Args: image: Input image in BGR format. resize_typ

High-level interface for one-click inference. Orchestrates the pipeline: pre_process -> forward -> post_process. Args:

Run the complete EfficientFormerV2 inference pipeline on one image. Args: image: Input image in BGR format.

Run the complete RepViT inference pipeline on one image. Args: image: Input image in BGR format. resize_type

Run the complete MobileNetV4 inference pipeline on one image. Args: image: Input image in BGR format. resize

Run the complete GoogLeNet inference pipeline on one image. Args: image: Input image in BGR format. resize_t

Run the complete matting pipeline on a single image. Args: image: Input image in BGR format. Returns: Alpha

Save the result image to disk.

Set optional runtime scheduling parameters. Args: priority: Scheduling priority in the range `0~255`. bpu_co

Apply the sigmoid activation function element-wise. Args: x: Input tensor in logit space. Returns: Tensor converted to proba

Build the fixed 32-slot text embedding tensor for selected prompts. Args: prompts: Prompt words that must exist in the o

Crop and perspective-rectify a text region from the image. Args: img (np.ndarray): Original image. box (np.ndarray):

Dilate contours using the ratio_prime parameter.

Compute minimum-area bounding rectangles from dilated polygons.

Perform class-wise rotated NMS. Args: rrects: Rotated rectangles in `(cx, cy, w, h, angle_rad)` format. scor

Execute recognition model inference on BPU. Args: input_tensor (np.ndarray): Float32 NCHW RGB tensor. Returns:

Run CTC decoding on recognition model output. Args: outputs (Dict[str, Dict[str, np.ndarray]]): Raw output tensors. Retu

Pre-process a cropped image for the recognition model. Args: cropped_img (np.ndarray): Cropped text region in BGR. Retur

Resize image with aspect-ratio-preserving scale and symmetric zero-padding. Args: image: Input image. target_size: Ta

Resolve the FCOS output order based on the original fixed protocol. Returns: Tuple[List[str], List[str], List[str]]: Cla

Compute IoU between two rotated rectangles. Args: a: First rotated rectangle in `(cx, cy, w, h, angle_rad)` format.

Map a rotated rectangle from input scale back to original image scale. Args: rrect (List[float]): Rotated rectangle on m

* @brief Convert BGR image to NV12 format */

* @brief Convert BGR image to NV12 format */

* @brief Convert BGR image to NV12 format * @param bgr_img Input BGR image * @return NV12 formatted image (height * 3/2, width, CV_8UC1) */

* @brief Convert BGR image to NV12 format */

递归遍历 image_root 目录，收集所有图像文件。 返回列表，每个元素为 (相对路径字符串, 绝对路径)。 相对路径使用 '/' 作为分隔符，并且不包含前导 './'。

Map class indices to BGR colors using the Cityscapes palette.

Composite the foreground (image) onto a background using the alpha matte. Args: image: Original image in BGR format. matte: Alpha

Replace the final Linear layer with a Conv2d 1x1 layer.

Crop instance masks using their corresponding bounding boxes. Args: masks: Predicted masks with shape ``(N, H, W)``. boxes: Bound

Crop instance masks using their corresponding bounding boxes. Args: masks (np.ndarray): Predicted masks with shape `(N, H, W)`.

Decode a single feature layer from the detection head. This function decodes the raw output tensor of one detection layer into bounding box p

Decode the raw model output into the final license plate string. Args: pred_data (np.ndarray): Raw model output after squeezing the

Decode one segmentation head output. Args: box_feat (np.ndarray): Bounding box branch output tensor. cls_feat (np.ndarray):

Decode one segmentation head output with DFL box regression. Unlike ``decode_seg_layer`` (which assumes pre-decoded 4-value bbox), this funct

Dequantize a quantized tensor to floating-point values. This function converts a quantized tensor (e.g., int8 or uint8) into floating-point v