update ema filter;

refix keypoint transform to orgimage

deploy/lite/include/keypoint_postprocess.h

@@ -33,7 +33,8 @@ void transform_preds(std::vector<float>& coords,
                      std::vector<float>& scale,
                      std::vector<int>& output_size,
                      std::vector<int>& dim,
-                     std::vector<float>& target_coords);
+                     std::vector<float>& target_coords,
+                     bool affine);
 void box_to_center_scale(std::vector<int>& box,
                          int width,
                          int height,

deploy/lite/src/keypoint_postprocess.cc

@@ -74,11 +74,26 @@ void transform_preds(std::vector<float>& coords,
                      std::vector<float>& scale,
                      std::vector<int>& output_size,
                      std::vector<int64_t>& dim,
-                     std::vector<float>& target_coords) {
-  cv::Mat trans(2, 3, CV_64FC1);
-  get_affine_transform(center, scale, 0, output_size, trans, 1);
-  for (int p = 0; p < dim[1]; ++p) {
-    affine_tranform(coords[p * 2], coords[p * 2 + 1], trans, target_coords, p);
+                     std::vector<float>& target_coords,
+                     bool affine=false) {
+  if (affine) {
+    cv::Mat trans(2, 3, CV_64FC1);
+    get_affine_transform(center, scale, 0, output_size, trans, 1);
+    for (int p = 0; p < dim[1]; ++p) {
+      affine_tranform(
+          coords[p * 2], coords[p * 2 + 1], trans, target_coords, p);
+    }
+  } else {
+    float heat_w = static_cast<float>(output_size[0]);
+    float heat_h = static_cast<float>(output_size[1]);
+    float x_scale = scale[0] / heat_w;
+    float y_scale = scale[1] / heat_h;
+    float offset_x = center[0] - scale[0] / 2.;
+    float offset_y = center[1] - scale[1] / 2.;
+    for (int i = 0; i < dim[1]; i++) {
+      target_coords[i * 3 + 1] = x_scale * coords[i * 2] + offset_x;
+      target_coords[i * 3 + 2] = y_scale * coords[i * 2 + 1] + offset_y;
+    }
   }
 }
 

deploy/python/det_keypoint_unite_infer.py

@@ -165,7 +165,7 @@ def topdown_unite_predict_video(detector,
             frame2, results, topdown_keypoint_detector, keypoint_batch_size,
             FLAGS.run_benchmark)
 
-        if FLAGS.smooth and len(keypoint_res['keypoint'][0])==1:
+        if FLAGS.smooth and len(keypoint_res['keypoint'][0]) == 1:
             current_keypoints = np.array(keypoint_res['keypoint'][0][0])
             smooth_keypoints = keypoint_smoothing.smooth_process(
                 current_keypoints)
@@ -215,7 +215,7 @@ class KeypointSmoothing(object):
                  fc_d=0.1,
                  fc_min=0.1,
                  beta=0.1,
-                 thres_mult=0.2):
+                 thres_mult=0.3):
         super(KeypointSmoothing, self).__init__()
         self.image_width = width
         self.image_height = height
@@ -234,7 +234,7 @@ class KeypointSmoothing(object):
         if self.filter_type == 'OneEuro':
             self.smooth_func = self.one_euro_filter
         elif self.filter_type == 'EMA':
-            self.smooth_func = self.exponential_smoothing
+            self.smooth_func = self.ema_filter
         else:
             raise ValueError('filter type must be one_euro or ema')
 
@@ -261,6 +261,7 @@ class KeypointSmoothing(object):
         else:
             result = self.smooth_func(current_keypoint, self.x_prev_hat[index],
                                       index)
+
         return result
 
     def one_euro_filter(self, x_cur, x_pre, index):
@@ -276,6 +277,11 @@ class KeypointSmoothing(object):
         self.x_prev_hat[index] = x_cur_hat
         return x_cur_hat
 
+    def ema_filter(self, x_cur, x_pre, index):
+        x_cur_hat = self.exponential_smoothing(x_cur, x_pre)
+        self.x_prev_hat[index] = x_cur_hat
+        return x_cur_hat
+
     def smoothing_factor(self, te, fc):
         r = 2 * math.pi * fc * te
         return r / (r + 1)