Fix regular detector obj default conf filter threshold (#570)

* Increase the all-classes default conf threshold val

* Add warnings to detector output objs config

* Remove heuristic for num_detected_classes

* Clear up comment

savant/deepstream/nvinfer/element_config.py

@@ -348,36 +348,48 @@ def process(self, msg, kwargs):
 
     # model type-specific parameters
     if issubclass(model_type, ObjectModel):
-        if not model_config.output.objects:
-            # try to load model object labels from file
-            label_file = model_config.get(
-                'label_file',
-                nvinfer_config['property'].get('labelfile-path')
-                if nvinfer_config
-                else None,
-            )
+        # model_config.output.objects is mandatory for object models
+        # but it may be autogenerated based on labelfile or num_detected_classes
+
+        label_file = model_config.get(
+            'label_file',
+            nvinfer_config['property'].get('labelfile-path')
+            if nvinfer_config
+            else None,
+        )
+        if model_config.output.objects:
+            # highest priority is using manually defined model_config.output.objects
             if label_file:
-                label_file_path = model_path / label_file
-                if label_file_path.is_file():
-                    with open(label_file_path.resolve(), encoding='utf8') as file_obj:
-                        model_config.output.objects = [
-                            NvInferObjectModelOutputObject(
-                                class_id=class_id, label=label
-                            )
-                            for class_id, label in enumerate(
-                                file_obj.read().splitlines()
-                            )
-                        ]
-                        model_config.output.num_detected_classes = len(
-                            model_config.output.objects
-                        )
-                    logger.info(
-                        'Model object labels have been loaded from "%s".',
-                        label_file_path,
-                    )
+                logger.warning(
+                    'Model output objects labels are defined manually '
+                    'and will be used instead of labels from "%s".',
+                    label_file,
+                )
 
-        # generate labels (enumerate)
-        if not model_config.output.objects and model_config.output.num_detected_classes:
+        elif label_file:
+            # try to load model object labels from file
+            label_file_path = model_path / label_file
+            if label_file_path.is_file():
+                with open(label_file_path.resolve(), encoding='utf8') as file_obj:
+                    model_config.output.objects = [
+                        NvInferObjectModelOutputObject(class_id=class_id, label=label)
+                        for class_id, label in enumerate(file_obj.read().splitlines())
+                    ]
+                if model_config.output.num_detected_classes:
+                    logger.warning(
+                        'Ignoring manually set value for '
+                        '(model_config.output.num_detected_classes) '
+                        'because labelfile is used.'
+                    )
+                model_config.output.num_detected_classes = len(
+                    model_config.output.objects
+                )
+                logger.info(
+                    'Model object labels have been loaded from "%s".',
+                    label_file_path,
+                )
+        elif model_config.output.num_detected_classes:
+            # generate labels (enumerate)
             model_config.output.objects = [
                 NvInferObjectModelOutputObject(class_id=class_id, label=str(class_id))
                 for class_id in range(model_config.output.num_detected_classes)

savant/deepstream/nvinfer/file_config.py

@@ -272,7 +272,7 @@ def from_model(
         # setup class-attrs for object model (detector)
         # set a high confidence threshold initially for all classes
         # to filter out only the desired classes
-        config['class-attrs-all'] = {'pre-cluster-threshold': 1.1}
+        config['class-attrs-all'] = {'pre-cluster-threshold': 1e10}
         # replace class-attrs parameters with selector kwargs
         for obj in model_config.output.objects:
             class_attrs = {}