diff --git a/blocky/CMakeLists.txt b/blocky/CMakeLists.txt
index e73b709..af0e8b0 100644
--- a/blocky/CMakeLists.txt
+++ b/blocky/CMakeLists.txt
@@ -7,6 +7,8 @@ target_sources(blocky
 
     bytes.hh
     common.hh
+    embed.hh
+    extract.hh
     video_encoder.hh
     video_decoder.hh
 )
diff --git a/blocky/embed.hh b/blocky/embed.hh
index af8712f..9bd9047 100644
--- a/blocky/embed.hh
+++ b/blocky/embed.hh
@@ -8,6 +8,8 @@
 #include "video_encoder.hh"
 
 
+namespace blky {
+
 void Embed(const std::vector<uint32_t>& features,
            const std::string& dst_path,
            const std::string& video_path,
@@ -68,21 +70,24 @@ void Embed(const std::vector<uint32_t>& features,
       feat_pix.resize(feat_size_x*feat_size_y);
       for (uint32_t y = 0; y < feat_size_y; ++y) {
         const uint32_t ay = y+feat_offset_y;
-        std::memcpy(feat_pix.data()+(y/2)*(feat_size_x/2), U.data()+(ay/2)*(w/2)+feat_offset_x/2, feat_size_x/2);
-        std::memcpy(feat_pix.data()+(y/2)*(feat_size_x/2)+feat_pix.size()/2, V.data()+(ay/2)*(w/2)+feat_offset_x/2, feat_size_x/2);
+        std::memcpy(feat_pix.data()+y*feat_size_x, Y.data()+ay*w+feat_offset_x, feat_size_x);
+        //std::memcpy(feat_pix.data()+(y/2)*(feat_size_x/2), U.data()+(ay/2)*(w/2)+feat_offset_x/2, feat_size_x/2);
+        //std::memcpy(feat_pix.data()+(y/2)*(feat_size_x/2)+feat_pix.size()/2, V.data()+(ay/2)*(w/2)+feat_offset_x/2, feat_size_x/2);
       }
       last_feat = feat;
+      if (enc) enc->ForceIntraFrame();
     }
     for (uint32_t y = 0; y < feat_size_y; ++y) {
       const uint32_t ay = y+feat_offset_y;
-      std::memcpy(U.data()+(ay/2)*(w/2)+feat_offset_x/2, feat_pix.data()+(y/2)*(feat_size_x/2), feat_size_x/2);
-      std::memcpy(V.data()+(ay/2)*(w/2)+feat_offset_x/2, feat_pix.data()+(y/2)*(feat_size_x/2)+feat_pix.size()/2, feat_size_x/2);
+      std::memcpy(Y.data()+ay*w+feat_offset_x, feat_pix.data()+y*feat_size_x, feat_size_x);
+      //std::memcpy(U.data()+(ay/2)*(w/2)+feat_offset_x/2, feat_pix.data()+(y/2)*(feat_size_x/2), feat_size_x/2);
+      //std::memcpy(V.data()+(ay/2)*(w/2)+feat_offset_x/2, feat_pix.data()+(y/2)*(feat_size_x/2)+feat_pix.size()/2, feat_size_x/2);
     }
 
     // create an encoder if not yet
     if (!enc) {
       SEncParamBase param = {};
-      param.iUsageType     = SCREEN_CONTENT_REAL_TIME;
+      param.iUsageType     = CAMERA_VIDEO_REAL_TIME;
       param.iPicWidth      = static_cast<int>(w);
       param.iPicHeight     = static_cast<int>(h);
       param.fMaxFrameRate  = 30;
@@ -105,3 +110,5 @@ void Embed(const std::vector<uint32_t>& features,
     enc->Encode(dst);
   }
 }
+
+}  // namespace blky
diff --git a/blocky/extract.hh b/blocky/extract.hh
new file mode 100644
index 0000000..bb3d186
--- /dev/null
+++ b/blocky/extract.hh
@@ -0,0 +1,65 @@
+#pragma once
+
+#include <cstdint>
+#include <cstring>
+#include <tuple>
+#include <vector>
+
+#include "common.hh"
+#include "video_decoder.hh"
+
+
+namespace blky {
+
+std::vector<double> Extract(
+    VideoDecoder& dec,
+    std::tuple<uint32_t, uint32_t> block_num,
+    std::tuple<uint32_t, uint32_t> sensor_num,
+    uint32_t utime) {
+  std::vector<double> ret;
+
+  blky_extractor_t ex = {};
+  ex.block_num_x        = std::get<0>(block_num);
+  ex.block_num_y        = std::get<1>(block_num);
+  ex.sensor_num_block_x = std::get<0>(sensor_num);
+  ex.sensor_num_block_y = std::get<1>(sensor_num);
+  ex.samples_per_pix    = 1;
+  ex.pix_stride         = 1;
+  ex.utime              = utime;
+  ex.correl_max_var     = 0.3;
+  ex.correl_min_avg     = 0.8;
+
+  blky_extractor_init(&ex);
+  while (dec.Decode()) {
+    const auto& frame = dec.frame();
+    if (frame.iBufferStatus != 1) {
+      continue;
+    }
+
+    const uint32_t w = static_cast<uint32_t>(frame.UsrData.sSystemBuffer.iWidth);
+    const uint32_t h = static_cast<uint32_t>(frame.UsrData.sSystemBuffer.iHeight);
+
+    const uint32_t stride_y  = static_cast<uint32_t>(frame.UsrData.sSystemBuffer.iStride[0]);
+    //const uint32_t stride_uv = static_cast<uint32_t>(frame.UsrData.sSystemBuffer.iStride[1]);
+
+    const uint8_t* const* srcp = frame.pDst;
+
+    const double wf = static_cast<double>(w) / static_cast<double>(stride_y);
+    const double verts[8] = {
+      0, 0,
+      0, 1,
+      wf, 1,
+      wf, 0,
+    };
+    if (blky_extractor_feed(&ex, srcp[0], stride_y, h, verts)) {
+      const size_t psize = ret.size();
+      ret.resize(ret.size() + ex.block_num);
+      std::memcpy(ret.data() + psize, ex.probs, ex.block_num * sizeof(ret[0]));
+    }
+  }
+  blky_extractor_deinit(&ex);
+
+  return ret;
+}
+
+}  // namespace blky
diff --git a/blocky/main.cc b/blocky/main.cc
index 37259d9..afd544d 100644
--- a/blocky/main.cc
+++ b/blocky/main.cc
@@ -8,6 +8,7 @@
 #include "common.hh"
 #include "bytes.hh"
 #include "embed.hh"
+#include "extract.hh"
 #include "features.hh"
 
 #include <args.hxx>
@@ -49,7 +50,7 @@ args::ValueFlag<std::tuple<uint32_t, uint32_t>> param_block_num {
   "int>0,int>0",
   "number of features",
   {"feature-num"},
-  {16, 16}
+  {8, 8}
 };
 args::ValueFlag<uint32_t> param_block_first {
   param_group,
@@ -85,6 +86,13 @@ args::ValueFlag<uint32_t> param_utime {
   {"utime"},
   10
 };
+args::ValueFlag<std::tuple<uint32_t, uint32_t>> param_sensor_num {
+  param_group,
+  "int>0,int>0",
+  "number of sensors (for extraction)",
+  {"sensor-num"},
+  {16, 16}
+};
 
 args::Group probgen_group {
   parser, "params for feature probability generator", args::Group::Validators::DontCare
@@ -217,8 +225,15 @@ try {
     switch (args::get(from)) {
     case kVideo:
       if (args::get(to) == kVideo) break;
-      // TODO extract feature probs // features = XX
-      assert(false);
+      if (!decoder) {
+        throw std::runtime_error {"decoder is empty"};
+      }
+      feature_probs = Extract(
+          *decoder,
+          args::get(param_block_num),
+          args::get(param_sensor_num),
+          args::get(param_utime));
+      /* fallthrough */
 
     case kFeatureProbs:
       if (args::get(to) == kFeatureProbs) break;
diff --git a/blocky/video_encoder.hh b/blocky/video_encoder.hh
index f855738..0645a79 100644
--- a/blocky/video_encoder.hh
+++ b/blocky/video_encoder.hh
@@ -58,6 +58,9 @@ class VideoEncoder final {
       }
     }
   }
+  void ForceIntraFrame() noexcept {
+    encoder_->ForceIntraFrame(true);
+  }
 
  private:
   std::ofstream file_;
diff --git a/liblocky/block.c b/liblocky/block.c
index ee6b80b..a2def36 100644
--- a/liblocky/block.c
+++ b/liblocky/block.c
@@ -18,7 +18,13 @@ double blky_block_estimate(
   var /= (double) n;
 
   // FIXME: calculate probability
-  if (var > max_var) return 0;
-  if (avg < min_avg) return 0;
-  return 1;
+  // if (var > max_var) return 0;
+  // if (avg < min_avg) return 0;
+  // return 1;
+  //
+  //if (var > 0.001) return 0;
+  (void) max_var;
+  (void) min_avg;
+  if (var > 0.01) return 0;
+  return blky_clamp(avg, 0, 1);
 }
diff --git a/liblocky/extractor.c b/liblocky/extractor.c
index 531e0e6..e8a0738 100644
--- a/liblocky/extractor.c
+++ b/liblocky/extractor.c
@@ -14,7 +14,7 @@ void blky_extractor_init(blky_extractor_t* ex) {
   assert(ex->utime > 0);
   assert(ex->utime <= BLKY_SENSOR_MAX_DUR);
   assert(ex->pix_stride > 0);
-  assert(ex->pix_stride > ex->samples_per_pix);
+  assert(ex->pix_stride >= ex->samples_per_pix);
   assert(ex->correl_max_var >= 0);
   assert(ex->correl_min_avg >= 0);
   assert(ex->correl_min_avg <= 1);