added support for both windows and unix

REQUIRE

@@ -4,3 +4,6 @@ Images 0.6
 Distributions 0.12
 FixedPointNumbers 0.3
 Compat 0.17
+FLANN
+NearestNeighbors
+Distances

src/ImageFeatures.jl

@@ -5,6 +5,8 @@ module ImageFeatures
 # package code goes here
 using Images, ColorTypes, FixedPointNumbers, Distributions
 using Compat
+using FLANN, Distances
+using NearestNeighbors
 
 include("core.jl")
 include("const.jl")

src/core.jl

@@ -94,13 +94,39 @@ function match_keypoints(keypoints_1::Keypoints, keypoints_2::Keypoints, desc_1,
         l_key = keypoints_1
         order = true
     end
-    hamming_distances = [hamming_distance(s, l) for s in smaller, l in larger]
+
     matches = Keypoints[]
-    for i in 1:length(smaller)
-        if any(hamming_distances[i, :] .< threshold)
-            id_min = indmin(hamming_distances[i, :])
-            push!(matches, order ? [l_key[id_min], s_key[i]] : [s_key[i], l_key[id_min]])
-            hamming_distances[:, id_min] = 1.0
+
+    ndims=length(larger[1])
+    n_large=length(larger)
+    n_small=length(smaller)
+
+    data=Matrix{Float64}(ndims, n_large);
+    for i in 1:ndims
+        for j in 1:n_large
+            data[i,j]=larger[j][i]?1:0
+        end
+    end
+
+    if is_windows() && Sys.WORD_SIZE==32
+        tree = KDTree(data, Cityblock())
+
+        for i in 1:n_small
+            idx, dist = NearestNeighbors.knn(tree, smaller[i], 1)
+            if dist[1]/ndims < threshold
+                id_min = idx[1]
+                push!(matches, order ? [l_key[id_min], s_key[i]] : [s_key[i], l_key[id_min]])
+            end
+        end   
+    else
+        tree = flann(data, FLANNParameters(), Cityblock())
+
+        for i in 1:n_small
+            idx, dist = FLANN.knn(tree, Vector{Float64}(smaller[i]), 1)
+            if dist[1]/ndims < threshold
+                id_min = idx[1]
+                push!(matches, order ? [l_key[id_min], s_key[i]] : [s_key[i], l_key[id_min]])
+            end
         end
     end
     matches