Merge "Store descriptor coefficients in 1 byte instead of 4"
diff --git a/y2020/vision/camera_reader.cc b/y2020/vision/camera_reader.cc
index c47e42b..f38a40e 100644
--- a/y2020/vision/camera_reader.cc
+++ b/y2020/vision/camera_reader.cc
@@ -189,12 +189,14 @@
// avoid crashes that only occur when specific features are matched.
CHECK(feature_table->has_field_location());
- const flatbuffers::Vector<float> *const descriptor =
+ const flatbuffers::Vector<uint8_t> *const descriptor =
feature_table->descriptor();
CHECK_EQ(descriptor->size(), 128u) << ": Unsupported feature size";
- cv::Mat(1, descriptor->size(), CV_32F,
- const_cast<void *>(static_cast<const void *>(descriptor->data())))
- .copyTo(features(cv::Range(i, i + 1), cv::Range(0, 128)));
+ const auto in_mat = cv::Mat(
+ 1, descriptor->size(), CV_8U,
+ const_cast<void *>(static_cast<const void *>(descriptor->data())));
+ const auto out_mat = features(cv::Range(i, i + 1), cv::Range(0, 128));
+ in_mat.convertTo(out_mat, CV_32F);
}
matcher_->add(features);
}
@@ -552,13 +554,21 @@
const cv::Mat &descriptors) {
const int number_features = keypoints.size();
CHECK_EQ(descriptors.rows, number_features);
+ CHECK_EQ(descriptors.cols, 128);
std::vector<flatbuffers::Offset<sift::Feature>> features_vector(
number_features);
for (int i = 0; i < number_features; ++i) {
- const auto submat = descriptors(cv::Range(i, i + 1), cv::Range(0, 128));
+ const auto submat =
+ descriptors(cv::Range(i, i + 1), cv::Range(0, descriptors.cols));
CHECK(submat.isContinuous());
- const auto descriptor_offset =
- fbb->CreateVector(reinterpret_cast<float *>(submat.data), 128);
+ flatbuffers::Offset<flatbuffers::Vector<uint8_t>> descriptor_offset;
+ {
+ uint8_t *data;
+ descriptor_offset = fbb->CreateUninitializedVector(128, &data);
+ submat.convertTo(
+ cv::Mat(1, descriptors.cols, CV_8U, static_cast<void *>(data)),
+ CV_8U);
+ }
sift::Feature::Builder feature_builder(*fbb);
feature_builder.add_descriptor(descriptor_offset);
feature_builder.add_x(keypoints[i].pt.x);
diff --git a/y2020/vision/sift/demo_sift_training.py b/y2020/vision/sift/demo_sift_training.py
index 3fa33cf..c78a44a 100644
--- a/y2020/vision/sift/demo_sift_training.py
+++ b/y2020/vision/sift/demo_sift_training.py
@@ -26,7 +26,8 @@
for keypoint, descriptor in zip(keypoints, descriptors):
Feature.FeatureStartDescriptorVector(fbb, len(descriptor))
for n in reversed(descriptor):
- fbb.PrependFloat32(n)
+ assert n == round(n)
+ fbb.PrependUint8(int(round(n)))
descriptor_vector = fbb.EndVector(len(descriptor))
Feature.FeatureStart(fbb)
diff --git a/y2020/vision/sift/sift.fbs b/y2020/vision/sift/sift.fbs
index 3e2daaf..24fd64d 100644
--- a/y2020/vision/sift/sift.fbs
+++ b/y2020/vision/sift/sift.fbs
@@ -9,7 +9,8 @@
// Represents a single feature extracted from an image.
table Feature {
- // Contains the descriptor data.
+ // Contains the descriptor data. OpenCV likes to represent them as floats, but
+ // they're really ubytes.
//
// TODO(Brian): These are scaled to be convertible to chars. Should we do
// that to minimize storage space? Or maybe int16?
@@ -17,7 +18,7 @@
// The size of this depends on the parameters. It is width*width*hist_bins.
// Currently we have width=4 and hist_bins=8, which results in a size of
// 4*4*8=128.
- descriptor:[float];
+ descriptor:[ubyte];
// Location of the keypoint.
x:float;
diff --git a/y2020/vision/tools/python_code/camera_param_test.cc b/y2020/vision/tools/python_code/camera_param_test.cc
index 5b959cd..4fdc2fe 100644
--- a/y2020/vision/tools/python_code/camera_param_test.cc
+++ b/y2020/vision/tools/python_code/camera_param_test.cc
@@ -143,12 +143,14 @@
cv::Mat features(training_image->features()->size(), 128, CV_32F);
for (size_t i = 0; i < training_image->features()->size(); ++i) {
const sift::Feature *feature_table = training_image->features()->Get(i);
- const flatbuffers::Vector<float> *const descriptor =
+ const flatbuffers::Vector<uint8_t> *const descriptor =
feature_table->descriptor();
CHECK_EQ(descriptor->size(), 128u) << ": Unsupported feature size";
- cv::Mat(1, descriptor->size(), CV_32F,
- const_cast<void *>(static_cast<const void *>(descriptor->data())))
- .copyTo(features(cv::Range(i, i + 1), cv::Range(0, 128)));
+ const auto in_mat = cv::Mat(
+ 1, descriptor->size(), CV_8U,
+ const_cast<void *>(static_cast<const void *>(descriptor->data())));
+ const auto out_mat = features(cv::Range(i, i + 1), cv::Range(0, 128));
+ in_mat.convertTo(out_mat, CV_32F);
cv::Point2f point_2d = Training2dPoint(train_image_index, i);
point_list_2d_.push_back(point_2d);
diff --git a/y2020/vision/tools/python_code/load_sift_training.py b/y2020/vision/tools/python_code/load_sift_training.py
index 6db3155..bf58e5f 100644
--- a/y2020/vision/tools/python_code/load_sift_training.py
+++ b/y2020/vision/tools/python_code/load_sift_training.py
@@ -71,7 +71,8 @@
# Build the Descriptor vector
Feature.FeatureStartDescriptorVector(fbb, len(descriptor))
for n in reversed(descriptor):
- fbb.PrependFloat32(n)
+ assert n == round(n)
+ fbb.PrependUint8(int(round(n)))
descriptor_vector = fbb.EndVector(len(descriptor))
# Add all the components to the each Feature