potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #10356 from dkurt:dnn_rfcn

Browse Source
This commit is contained in:
Vadim Pisarevsky
2017-12-20 14:19:46 +00:00
parent d2f13a8259 0ed2cbc931
commit b8a24b36ce
7 changed files with 1588 additions and 633 deletions
Download Patch File Download Diff File Expand all files Collapse all files
samples/dnn/faster_rcnn.cpp
+6 -61
View File
@@ -1,24 +1,3 @@
// Faster-RCNN models use custom layer called 'Proposal' written in Python. To
// map it into OpenCV's layer replace a layer node with [type: 'Python'] to the
// following definition:
// layer {
// name: 'proposal'
// type: 'Proposal'
// bottom: 'rpn_cls_prob_reshape'
// bottom: 'rpn_bbox_pred'
// bottom: 'im_info'
// top: 'rois'
// proposal_param {
// ratio: 0.5
// ratio: 1.0
// ratio: 2.0
// scale: 8
// scale: 16
// scale: 32
// }
// }
#include <iostream>
#include <opencv2/dnn.hpp>
#include <opencv2/dnn/all_layers.hpp>
#include <opencv2/imgproc.hpp>
@@ -50,9 +29,11 @@ int main(int argc, char** argv)
{
// Parse command line arguments.
CommandLineParser parser(argc, argv, keys);
parser.about( "This sample is used to run Faster-RCNN object detection with OpenCV.\n"
"You can get required models from https://github.com/rbgirshick/py-faster-rcnn" );
parser.about("This sample is used to run Faster-RCNN and R-FCN object detection "
"models with OpenCV. You can get required models from "
"https://github.com/rbgirshick/py-faster-rcnn (Faster-RCNN) and from "
"https://github.com/YuwenXiong/py-R-FCN (R-FCN). Corresponding .prototxt "
"files may be found at https://github.com/opencv/opencv_extra/tree/master/testdata/dnn.");
if (argc == 1 || parser.has("help"))
{
parser.printMessage();
@@ -68,19 +49,6 @@ int main(int argc, char** argv)
// Load a model.
Net net = readNetFromCaffe(protoPath, modelPath);
// Create a preprocessing layer that does final bounding boxes applying predicted
// deltas to objects locations proposals and doing non-maximum suppression over it.
LayerParams lp;
lp.set("code_type", "CENTER_SIZE"); // An every bounding box is [xmin, ymin, xmax, ymax]
lp.set("num_classes", 21);
lp.set("share_location", (int)false); // Separate predictions for different classes.
lp.set("background_label_id", 0);
lp.set("variance_encoded_in_target", (int)true);
lp.set("keep_top_k", 100);
lp.set("nms_threshold", 0.3);
lp.set("normalized_bbox", (int)false);
Ptr<Layer> detectionOutputLayer = DetectionOutputLayer::create(lp);
Mat img = imread(imagePath);
resize(img, img, Size(kInpWidth, kInpHeight));
Mat blob = blobFromImage(img, 1.0, Size(), Scalar(102.9801, 115.9465, 122.7717), false, false);
@@ -89,31 +57,8 @@ int main(int argc, char** argv)
net.setInput(blob, "data");
net.setInput(imInfo, "im_info");
std::vector<Mat> outs;
std::vector<String> outNames(3);
outNames[0] = "proposal";
outNames[1] = "bbox_pred";
outNames[2] = "cls_prob";
net.forward(outs, outNames);
Mat proposals = outs[0].colRange(1, 5).clone(); // Only last 4 columns.
Mat& deltas = outs[1];
Mat& scores = outs[2];
// Reshape proposals from Nx4 to 1x1xN*4
std::vector<int> shape(3, 1);
shape[2] = (int)proposals.total();
proposals = proposals.reshape(1, shape);
// Run postprocessing layer.
std::vector<Mat> layerInputs(3), layerOutputs(1), layerInternals;
layerInputs[0] = deltas.reshape(1, 1);
layerInputs[1] = scores.reshape(1, 1);
layerInputs[2] = proposals;
detectionOutputLayer->forward(layerInputs, layerOutputs, layerInternals);
// Draw detections.
Mat detections = layerOutputs[0];
Mat detections = net.forward();
const float* data = (float*)detections.data;
for (size_t i = 0; i < detections.total(); i += 7)
{