Semantic Segmentation

Very light weight semantic segmentation model.

Query image and prediction

BiWAKO.FastSCNN

Semantic Segmentation

Attributes:

Name	Type	Description
model_path	str	Path to the model file.
model	onnxruntime.InferenceSession	Inference session.
input_shape	tuple	Input shape of the model. Set to (384, 384) for fast_scnn384 and (1344, 768) for fast_scnn7681344.
input_name	str	Name of the input node.
output_name	str	Name of the output node.
mean	np.ndarray	Mean value of the dataset.
std	np.ndarray	Standard deviation of the dataset.
c_map		Color map for the semantic segmentation 19 object classes.

__init__(self, model='fast_scnn384', **kwargs) special

Initialize FastSCNN model.

Parameters:

Name	Type	Description	Default
model	str	Choice of model. Accept model name or path to the downloaded onnx file. If onnx file has not been downloaded, it will be downloaded automatically. Currently avaiable models are [fast_scnn384, fast_scnn7681344]. Defaults to "fast_scnn384".	'fast_scnn384'

_preprocess(self, image) private

Preprocess the image. Automatically called.

Preprocess

1. Resize to the input shape.
2. To RGB and normalize.
3. Add the mean and divide by the standard deviation.
4. Channel swap and float32.
5. Add batch dimension.

Parameters:

Name	Type	Description	Default
image	np.ndarray	Image to be preprocessed.	required

Returns:

Type	Description
np.ndarray	Preprocessed image.

predict(self, image)

Return the prediction map. The last channel has 19 classes.

Parameters:

Name	Type	Description	Default
image	Image	Image to be predicted. Accept path to the image or cv2 image.	required

Returns:

Type	Description
np.ndarray	Prediction map in the shape of (height, width, 19).

render(self, prediction, image, **kwargs)

Apply the prediction map to the image.

Parameters:

Name	Type	Description	Default
prediction	np.ndarray	Prediction map retuned by predict.	required
image	Image	Image to be rendered. Accept path to the image or cv2 image.	required

Returns:

Type	Description
np.ndarray	Rendered image.