AC: fix processing tf action detection (#3148)

* AC: fix processing tf action detection

* Update tools/accuracy_checker/openvino/tools/accuracy_checker/adapters/action_recognition.py

Co-authored-by: Anna Grebneva <[email protected]>

Co-authored-by: Anna Grebneva <[email protected]>

Author: eaidova

tools/accuracy_checker/openvino/tools/accuracy_checker/adapters/action_recognition.py

@@ -145,14 +145,17 @@ def process(self, raw, identifiers, frame_meta):
         raw_outputs = self._extract_predictions(raw, frame_meta)
         if not self.outputs_verified:
             self._get_output_names(raw_outputs)
+        input_shape = list(frame_meta[0].get('input_shape', {'data': (1, 3, 416, 416)}).values())[0]
+        nchw_layout = input_shape[1] == 3
         prior_boxes = raw_outputs[self.priorbox_out][0][0].reshape(-1, 4) if not self.multihead else None
         prior_variances = raw_outputs[self.priorbox_out][0][1].reshape(-1, 4) if not self.multihead else None
 
-        head_shifts = self.estimate_head_shifts(raw_outputs, self. head_sizes, self.add_conf_outs, self.multihead)
+        head_shifts = self.estimate_head_shifts(
+            raw_outputs, self. head_sizes, self.add_conf_outs, self.multihead, nchw_layout)
 
         for batch_id, identifier in enumerate(identifiers):
             labels, class_scores, x_mins, y_mins, x_maxs, y_maxs, main_scores = self.prepare_detection_for_id(
-                batch_id, raw_outputs, prior_boxes, prior_variances, head_shifts
+                batch_id, raw_outputs, prior_boxes, prior_variances, head_shifts, nchw=nchw_layout
             )
             action_prediction = ActionDetectionPrediction(
                 identifier, labels, class_scores, main_scores, x_mins, y_mins, x_maxs, y_maxs
@@ -167,12 +170,15 @@ def process(self, raw, identifiers, frame_meta):
         return result
 
     def prepare_detection_for_id(self, batch_id, raw_outputs, prior_boxes, prior_variances, head_shifts,
-                                 default_label=0):
+                                 default_label=0, nchw=True):
         num_detections = raw_outputs[self.loc_out][batch_id].size // 4
         locs = raw_outputs[self.loc_out][batch_id].reshape(-1, 4)
         main_conf = raw_outputs[self.main_conf_out][batch_id].reshape(num_detections, -1)
 
         add_confs = [raw_outputs[layer][batch_id] for layer in self.add_conf_outs]
+        if not nchw:
+            add_confs = [np.transpose(conf_l, (2, 0, 1)) for conf_l in add_confs]
+
         if self.multihead:
             spatial_sizes = [layer.shape[1:] for layer in add_confs]
             add_confs = [layer.reshape(self.num_action_classes, -1) for layer in add_confs]
@@ -238,14 +244,16 @@ def decode_box(prior, var, deltas):
         return decoded_xmin, decoded_ymin, decoded_xmax, decoded_ymax
 
     @staticmethod
-    def estimate_head_shifts(raw_outputs, head_sizes, add_conf_outs, multihead_net):
+    def estimate_head_shifts(raw_outputs, head_sizes, add_conf_outs, multihead_net, nchw=True):
         layer_id = 0
         head_shift = 0
         head_shifts = [0]
         for head_size in head_sizes:
             for _ in range(head_size):
                 layer = add_conf_outs[layer_id]
                 layer_shape = raw_outputs[layer][0].shape
+                if len(layer_shape) == 3 and not nchw:
+                    layer_shape = layer_shape[::-1]
                 layer_size = np.prod(layer_shape[1:]) if multihead_net else np.prod(layer_shape[:2])
                 head_shift += layer_size
                 layer_id += 1
@@ -297,19 +305,17 @@ def find_layer(regex, output_name, all_outputs):
 
         self.loc_out = find_layer(loc_out_regex, 'loc', raw_outputs)
         self.main_conf_out = find_layer(main_conf_out_regex, 'main confidence', raw_outputs)
-        self.priorbox_out = self.check_output_name(self.priorbox_out, raw_outputs)
+        if hasattr(self, 'priorbox_out'):
+            self.priorbox_out = self.check_output_name(self.priorbox_out, raw_outputs)
         self.outputs_verified = True
-        if contains_all(raw_outputs, self.add_conf_outs):
-            return
-        add_conf_result = [layer_name + '/sink_port_0' for layer_name in self.add_conf_outs]
-        if contains_all(raw_outputs, add_conf_result):
-            self.add_conf_outs = add_conf_result
+        add_conf_outs = [self.check_output_name(layer, raw_outputs) for layer in self.add_conf_outs]
+        if contains_all(add_conf_outs):
+            self.add_conf_outs = add_conf_outs
             return
-        add_conf_with_bias = [layer_name + '/add_' for layer_name in self.add_conf_outs]
+
+        add_conf_with_bias = [self.check_output_name(layer_name + '/add_', raw_outputs)
+                              for layer_name in self.add_conf_outs]
         if contains_all(raw_outputs, add_conf_with_bias):
             self.add_conf_outs = add_conf_with_bias
             return
-        add_conf_with_bias_result = [layer_name + '/add_/sink_port_0' for layer_name in self.add_conf_outs]
-        if contains_all(raw_outputs, add_conf_with_bias_result):
-            self.add_conf_outs = add_conf_with_bias_result
         return