Table Of Contents
The ProposalLayer
plugin generates the first-stage detection (ROI candidates) out of the scores, refinement info from RPN (Region Proposal Network) and pre-defined anchors. It is used in sampleUffMaskRCNN
.
This plugin supports the NCHW format. It takes two input tenosrs: object_score
and object_delta
object_score
is the objectness score from RPN. object_score
's shape is [N, anchors, 2, 1]
where N
is the batch_size, anchors
is the total number of anchors and 2
means 2 classes of objectness - foreground and background.
object_delta
is the refinement info from RPN of shape [N, anchors, 4, 1]
. 4
refers to the 4 elements of refinement information - [dy, dx, dh, dw]
This plugin generates one output tensor of shape [N, keep_topk, 4]
where keep_topk
is the maximum number of detections left after NMS and 4
refers to coordinates of ROI candidates [y1, x1, y2, x2]
The default anchors are generated in this plugin during initialization
method. For ResNet101 with 1024*1024 input image, the number of anchors can be computed as:
This plugin has the plugin creator class ProposalLayerPluginCreator
and the plugin class ProposalLayer
.
The following parameters were used to create ProposalLayer
instance:
Type | Parameter | Description |
---|---|---|
int | prenms_topk | The number of ROIs which will be kept before NMS. |
int | keep_topk | Number of detections will be kept after NMS. |
float | iou_threshold | IOU threshold value used in NMS. |
int[3] | image_size | Input image size in CHW. Only supports C=3 and defaults to [3,1024,1024]. |
The following resources provide a deeper understanding of the ProposalLayer
plugin:
For terms and conditions for use, reproduction, and distribution, see the TensorRT Software License Agreement documentation.
July 2020: Add (optional) plugin parameter for specifying image size. June 2019: First release of proposeLayerPlugin.
There are no known issues in this plugin.