filter DSPSIFT keypoints based on border size

OpenDroneMap · Jul 24, 2024 · 1a5f417 · 1a5f417
1 parent 7406fb9
commit 1a5f417
Show file tree

Hide file tree

Showing 4 changed files with 18 additions and 5 deletions.
diff --git a/opensfm/features.py b/opensfm/features.py
@@ -656,8 +656,6 @@ def extract_features(
 
     xs = points[:, 0].round().astype(int)
     ys = points[:, 1].round().astype(int)
-    xs = np.clip(xs, 0, image.shape[1] - 1)
-    ys = np.clip(ys, 0, image.shape[0] - 1)
     colors = image[ys, xs]
     if image.shape[2] == 1:
         colors = np.repeat(colors, 3).reshape((-1, 3))

diff --git a/opensfm/src/features/dspsift.h b/opensfm/src/features/dspsift.h
@@ -6,6 +6,6 @@ namespace features {
 
 py::tuple dspsift(foundation::pyarray_f image, float peak_threshold,
                 float edge_threshold, int target_num_features,
-                bool feature_root, bool domain_size_pooling, bool estimate_affine_shape);
+                bool feature_root, bool domain_size_pooling, bool estimate_affine_shape, int border_size);
 
 }
diff --git a/opensfm/src/features/python/pybind.cc b/opensfm/src/features/python/pybind.cc
@@ -62,7 +62,8 @@ PYBIND11_MODULE(pyfeatures, m) {
         py::arg("target_num_features") = 0,
         py::arg("feature_root") = true,
         py::arg("domain_size_pooling") = true,
-        py::arg("estimate_affine_shape") = true);
+        py::arg("estimate_affine_shape") = true,
+        py::arg("border_size") = 5);
 
   m.def("match_using_words", features::match_using_words);
   m.def("compute_vlad_descriptor", features::compute_vlad_descriptor,

diff --git a/opensfm/src/features/src/dspsift.cc b/opensfm/src/features/src/dspsift.cc
@@ -8,6 +8,7 @@ Copyright (c) 2023, ETH Zurich and UNC Chapel Hill.
 #include <iostream>
 #include <vector>
 #include <memory>
+#include <algorithm>
 #include <Eigen/Core>
 
 extern "C" {
@@ -47,11 +48,18 @@ namespace features {
 
 py::tuple dspsift(foundation::pyarray_f image, float peak_threshold,
                 float edge_threshold, int target_num_features, 
-                bool feature_root, bool domain_size_pooling, bool estimate_affine_shape) {
+                bool feature_root, bool domain_size_pooling, bool estimate_affine_shape, int border_size) {
   if (!image.size()) {
     return py::none();
   }
 
+  int width = image.shape(1);
+  int height = image.shape(0);
+
+  if (border_size*2 >= width || border_size*2 >= height) {
+    return py::none();
+  }
+
   FeatureDescriptors descriptors;
   FeatureKeypoints keypoints;
   int keypoints_count = 0;
@@ -103,7 +111,13 @@ py::tuple dspsift(foundation::pyarray_f image, float peak_threshold,
     }
 
     VlCovDetFeature* features = vl_covdet_get_features(covdet.get());
+
+    VlCovDetFeature* middle = std::partition(features, features + num_features, [border_size, width, height](const VlCovDetFeature& feature) {
+      return feature.frame.x >= border_size && feature.frame.x < width - border_size &&
+             feature.frame.y >= border_size && feature.frame.y < height - border_size;
+    });
 
+    num_features = middle - features;
 
     // Sort features according to detected octave and scale.
     std::sort(