List of Custom Operators

Computation operators

Proposal

Obtains the accurate proposals in Faster R-CNN, based on the foreground of rpn_cls_prob and the refined anchors obtained through bounding box regression of rpn_bbox_pred.

ROIPooling

Performs region-of-interest pooling in Faster R-CNN, which is mainly used for an object detection task.

PSROIPooling

Performs position-sensitive region-of-interest pooling in R-FCN, which is mainly used for an object detection task.

Computation operators

Reverse

Reverses the dimensions of a tensor.

Upsample

Performs upsampling using pooling mask.

Used in the YOLO network.

Normalize

Normalizes the input tensor along the channel dimension using an L2 norm in SSD.

Reorg

Rearranges blocks of spatial data into depth, or vice versa, in Darknet.

Implemented as a PassThrough operator as defined in the operator specifications.

ROIAlign

Obtains the range of interest (ROI) feature matrix from the feature map.

ShuffleChannel

Groups and permutes data in the channel dimension of the input.

YOLO (YOLO/Detection/Region)

Generates coordinates, confidence scores, and category probability of the BBoxes on the feature map output by the convolutional network. All the YOLO, Detection, and Region operators need to be replaced by the YOLO operator.

PriorBox

Generates prior boxes based on the input parameters in the SSD network.

SpatialTransformer

Performs affine transformation.

Postprocessing operators

YoloV3DetectionOutput

Generates coordinates, confidence scores, and category probability of the BBoxes on the feature map output by the convolutional network. This operator is used for YOLOv3 postprocessing.

YoloV2DetectionOutput

Generates coordinates, confidence scores, and category probability of the BBoxes on the feature map output by the convolutional network. This operator is used for YOLOv2 postprocessing.

SSDDetectionOutput

Integrates the BBoxes, BBox offsets, and scores, and outputs object predictions of SSD. This operator is used for SSD network postprocessing.

FSRDetectionOutput

Classifies the results, and outputs the final number, coordinates, category probability, and category indexes of BBoxes. This operator is used for Faster R-CNN postprocessing.