potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-31 00:03:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #12243 from dkurt:dnn_tf_mask_rcnn

Browse Source 
* Support Mask-RCNN from TensorFlow

* Fix a sample
This commit is contained in:
Dmitry Kurtaev
2018-08-24 14:47:32 +03:00
committed by Alexander Alekhin
parent 4f360f8b1a
commit 472b71ecef
9 changed files with 600 additions and 153 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/crop_and_resize_layer.cpp
+7
View File
@@ -99,6 +99,13 @@ public:
}
}
}
if (boxes.rows < out.size[0])
{
// left = top = right = bottom = 0
std::vector<cv::Range> dstRanges(4, Range::all());
dstRanges[0] = Range(boxes.rows, out.size[0]);
out(dstRanges).setTo(inp.ptr<float>(0, 0, 0)[0]);
}
}
private:
modules/dnn/src/layers/detection_output_layer.cpp
+40 -10
View File
@@ -115,6 +115,7 @@ public:
// It's true whenever predicted bounding boxes and proposals are normalized to [0, 1].
bool _bboxesNormalized;
bool _clip;
bool _groupByClasses;
enum { _numAxes = 4 };
static const std::string _layerName;
@@ -183,6 +184,7 @@ public:
_locPredTransposed = getParameter<bool>(params, "loc_pred_transposed", 0, false, false);
_bboxesNormalized = getParameter<bool>(params, "normalized_bbox", 0, false, true);
_clip = getParameter<bool>(params, "clip", 0, false, false);
_groupByClasses = getParameter<bool>(params, "group_by_classes", 0, false, true);
getCodeType(params);
@@ -381,7 +383,7 @@ public:
{
count += outputDetections_(i, &outputsData[count * 7],
allDecodedBBoxes[i], allConfidenceScores[i],
allIndices[i]);
allIndices[i], _groupByClasses);
}
CV_Assert(count == numKept);
}
@@ -497,7 +499,7 @@ public:
{
count += outputDetections_(i, &outputsData[count * 7],
allDecodedBBoxes[i], allConfidenceScores[i],
allIndices[i]);
allIndices[i], _groupByClasses);
}
CV_Assert(count == numKept);
}
@@ -505,9 +507,36 @@ public:
size_t outputDetections_(
const int i, float* outputsData,
const LabelBBox& decodeBBoxes, Mat& confidenceScores,
const std::map<int, std::vector<int> >& indicesMap
const std::map<int, std::vector<int> >& indicesMap,
bool groupByClasses
)
{
std::vector<int> dstIndices;
std::vector<std::pair<float, int> > allScores;
for (std::map<int, std::vector<int> >::const_iterator it = indicesMap.begin(); it != indicesMap.end(); ++it)
{
int label = it->first;
if (confidenceScores.rows <= label)
CV_Error_(cv::Error::StsError, ("Could not find confidence predictions for label %d", label));
const std::vector<float>& scores = confidenceScores.row(label);
const std::vector<int>& indices = it->second;
const int numAllScores = allScores.size();
allScores.reserve(numAllScores + indices.size());
for (size_t j = 0; j < indices.size(); ++j)
{
allScores.push_back(std::make_pair(scores[indices[j]], numAllScores + j));
}
}
if (!groupByClasses)
std::sort(allScores.begin(), allScores.end(), util::SortScorePairDescend<int>);
dstIndices.resize(allScores.size());
for (size_t j = 0; j < dstIndices.size(); ++j)
{
dstIndices[allScores[j].second] = j;
}
size_t count = 0;
for (std::map<int, std::vector<int> >::const_iterator it = indicesMap.begin(); it != indicesMap.end(); ++it)
{
@@ -524,14 +553,15 @@ public:
for (size_t j = 0; j < indices.size(); ++j, ++count)
{
int idx = indices[j];
int dstIdx = dstIndices[count];
const util::NormalizedBBox& decode_bbox = label_bboxes->second[idx];
outputsData[count * 7] = i;
outputsData[count * 7 + 1] = label;
outputsData[count * 7 + 2] = scores[idx];
outputsData[count * 7 + 3] = decode_bbox.xmin;
outputsData[count * 7 + 4] = decode_bbox.ymin;
outputsData[count * 7 + 5] = decode_bbox.xmax;
outputsData[count * 7 + 6] = decode_bbox.ymax;
outputsData[dstIdx * 7] = i;
outputsData[dstIdx * 7 + 1] = label;
outputsData[dstIdx * 7 + 2] = scores[idx];
outputsData[dstIdx * 7 + 3] = decode_bbox.xmin;
outputsData[dstIdx * 7 + 4] = decode_bbox.ymin;
outputsData[dstIdx * 7 + 5] = decode_bbox.xmax;
outputsData[dstIdx * 7 + 6] = decode_bbox.ymax;
}
}
return count;
modules/dnn/src/layers/resize_layer.cpp
+11 -5
View File
@@ -33,9 +33,7 @@ public:
interpolation = params.get<String>("interpolation");
CV_Assert(interpolation == "nearest" || interpolation == "bilinear");
bool alignCorners = params.get<bool>("align_corners", false);
if (alignCorners)
CV_Error(Error::StsNotImplemented, "Resize with align_corners=true is not implemented");
alignCorners = params.get<bool>("align_corners", false);
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
@@ -66,8 +64,15 @@ public:
outHeight = outputs[0].size[2];
outWidth = outputs[0].size[3];
}
scaleHeight = static_cast<float>(inputs[0]->size[2]) / outHeight;
scaleWidth = static_cast<float>(inputs[0]->size[3]) / outWidth;
if (alignCorners && outHeight > 1)
scaleHeight = static_cast<float>(inputs[0]->size[2] - 1) / (outHeight - 1);
else
scaleHeight = static_cast<float>(inputs[0]->size[2]) / outHeight;
if (alignCorners && outWidth > 1)
scaleWidth = static_cast<float>(inputs[0]->size[3] - 1) / (outWidth - 1);
else
scaleWidth = static_cast<float>(inputs[0]->size[3]) / outWidth;
}
void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
@@ -166,6 +171,7 @@ protected:
int outWidth, outHeight, zoomFactorWidth, zoomFactorHeight;
String interpolation;
float scaleWidth, scaleHeight;
bool alignCorners;
};
modules/dnn/test/test_tf_importer.cpp
+52
View File
@@ -537,4 +537,56 @@ TEST(Test_TensorFlow, two_inputs)
normAssert(out, firstInput + secondInput);
}
TEST(Test_TensorFlow, Mask_RCNN)
{
std::string proto = findDataFile("dnn/mask_rcnn_inception_v2_coco_2018_01_28.pbtxt", false);
std::string model = findDataFile("dnn/mask_rcnn_inception_v2_coco_2018_01_28.pb", false);
Net net = readNetFromTensorflow(model, proto);
Mat img = imread(findDataFile("dnn/street.png", false));
Mat refDetections = blobFromNPY(path("mask_rcnn_inception_v2_coco_2018_01_28.detection_out.npy"));
Mat refMasks = blobFromNPY(path("mask_rcnn_inception_v2_coco_2018_01_28.detection_masks.npy"));
Mat blob = blobFromImage(img, 1.0f, Size(800, 800), Scalar(), true, false);
net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.setInput(blob);
// Mask-RCNN predicts bounding boxes and segmentation masks.
std::vector<String> outNames(2);
outNames[0] = "detection_out_final";
outNames[1] = "detection_masks";
std::vector<Mat> outs;
net.forward(outs, outNames);
Mat outDetections = outs[0];
Mat outMasks = outs[1];
normAssertDetections(refDetections, outDetections, "", /*threshold for zero confidence*/1e-5);
// Output size of masks is NxCxHxW where
// N - number of detected boxes
// C - number of classes (excluding background)
// HxW - segmentation shape
const int numDetections = outDetections.size[2];
int masksSize[] = {1, numDetections, outMasks.size[2], outMasks.size[3]};
Mat masks(4, &masksSize[0], CV_32F);
std::vector<cv::Range> srcRanges(4, cv::Range::all());
std::vector<cv::Range> dstRanges(4, cv::Range::all());
outDetections = outDetections.reshape(1, outDetections.total() / 7);
for (int i = 0; i < numDetections; ++i)
{
// Get a class id for this bounding box and copy mask only for that class.
int classId = static_cast<int>(outDetections.at<float>(i, 1));
srcRanges[0] = dstRanges[1] = cv::Range(i, i + 1);
srcRanges[1] = cv::Range(classId, classId + 1);
outMasks(srcRanges).copyTo(masks(dstRanges));
}
cv::Range topRefMasks[] = {Range::all(), Range(0, numDetections), Range::all(), Range::all()};
normAssert(masks, refMasks(&topRefMasks[0]));
}
}