cropAndResize Plugin

Table Of Contents

• Description
  • Structure
• Parameters
• Additional resources
• License
• Changelog
• Known issues

Description

NOTE: The IPluginV2Ext version of this plugin is deprecated since TensorRT 9.0. IPluginV2DynamicExt-based CropAndResizeDynamic is recommended instead.

The cropAndResizePlugin performs object detection for the Faster R-CNN model. This plugin is included in TensorRT.

cropAndResizePlugin implements the TensorFlow style of ROIPooling(a.k.a. CropAndResize). It crops multiple region of interests(ROIs) from the input image with given ROI coordinates and then (bilinearly) resizes the cropped patches to a target spatial(width and height) size.

Note this implementation is different from the original Caffe implement of ROIPooling. Also, this implementation doesn't fuse the ROIPooling layer and the Proposal layer into a single layer as the nvFasterRCNN plugin does.

This plugin is optimized for the above steps and it allows you to do Faster R-CNN inference in TensorRT.

Structure

The cropAndResizePlugin takes two inputs; feature_maps and rois.

feature_maps Input featuremaps where the patches are going to be cropped from. feature_maps has shape [N, C, H, W] where N is the batch size, C is the channel number, H is the height, W is the width. The feature map contains the features that we want to crop patches from and do classification after that to determine the class of each patch(potential object) and the bounding boxes deltas for that object.

rois Coordinates of region of interest bounding boxes, normalized to the range of (0, 1). rois has shape [N, B, 4, 1], where N is the batch size, B is the maximum number of region of interest bounding boxes per image(featuremap), and 4 represents and region of interest bounding box coordinates [y_1, x_1, y_2, x_2]. Here, x_1 and y_1 are the coordinates of bounding box at the top-left corner, and x_2 and y_2 are the coordinates of bounding box at the bottom-right corner.

The cropAndResizePlugin generates the following one outputs:

pfmap ROI pooled feature map corresponding to the region of interest. pfmap has shape [N, B, C, crop_height, crop_width] where N is the batch size, B is the maximum number of region of interest bounding boxes, C is the number of channels in the feature map, crop_height is the height of ROI pooled feature map, and crop_width is the width of ROI pooled feature map.

The ROI pooling step uses the inferred region of interest bounding boxes information to extract its corresponding regions on feature map, and does POI pooling to get uniformly shaped features from different shaped region of interest bounding boxes.

Parameters

cropAndResizePlugin has plugin creator class cropAndResizePluginCreator and plugin class CropAndResizePlugin.

The parameters are defined below and consists of the following attributes:

Type	Parameter	Description
int	crop_width	The height of the output in pixels after ROI pooling on the feature map.
int	crop_height	The width of the output in pixels after ROI pooling on the feature map.

Additional resources

The following resources provide a deeper understanding of the cropAndResizePlugin plugin:

Networks:

• Faster R-CNN

Documentation:

• Original ROI Pooling Definition from Fast R-CNN
• CropAndResize Op in TensorFlow
• tf.image.crop_and_resize API

License

For terms and conditions for use, reproduction, and distribution, see the TensorRT Software License Agreement documentation.

Changelog

June 2023 Add deprecation note for IPluginV2Ext version of the plugin.

May 2019 This is the first release of this README.md file.

Known issues

There are no known issues in this plugin.