BatchMultiClassNMS

Description

Computes the NMS of the input tensor boxes and scores. It supports multiple batches and classes, suppresses non-maximum elements, and retains maximum elements.

Input

• boxes: input tensor. The shape is [batch, num_anchors, num_classes, 4]. batch indicates the batch size of images, num_anchors indicates the number of boxes, num_classes indicates the number of detection classes, and 4 indicates x0, x1, y0, and y1. The data type can be float16 or float32.
• scores: input tensor, which indicates the box. The shape is [batch, num_anchors, num_classes]. The data type can be float16 or float32.
• clip_window (optional): input tensor, indicating the window size. The shape is [batch, 4]. The data type can be float16 or float32.
• num_valid_boxes (optional): input tensor, indicating the number of valid boxes in each batch. The data type can be int32.

Attribute

• score_threshold: float, indicating the score filter.
• iou_threshold: float, indicating the IOU overlap threshold.
• max_size_per_class: int, specifying the number of NMS outputs for each class.
• max_total_size: int, specifying the number of NMS outputs for each batch.
• change_coordinate_frame (optional): bool, indicating whether to normalize coordinates after cropping.
• transpose_box (optional): bool, indicating whether to insert transpose before this operation. The value must be false.

Output

• nmsed_boxes: output tensor, specifying the output NMS boxes for each batch. The data type can be float16 or float32.
• nmsed_scores: output tensor, specifying the output NMS scores for each batch. The data type can be float16 or float32.
• nmsed_classes: output tensor, specifying the output NMS classes for each batch. The data type can be float16 or float32.
• nmsed_num: output tensor, specifying the valid number of nmsed_boxes. The data type can be int32.

Constraints

None.

ONNX Opset Support

Opset v8/v9/v10/v11/v12/v13/v14/v15/v16