pythia.utils.maskrcnn#

Simple python MASKRCNN output parser.

The function extract_maskrcnn_mask should be used.

The core function, resize_mask_vec, is based on its cpp counterpart resizeMask, from deepstream cpp sources. It was ported to python and vectorized via numpy to improve speed.

Author: <Pablo Woolvett pablowoolvett@gmail.com>

class pythia.utils.maskrcnn.DsBBox(*args, **kwargs)[source]#

Bases: dict

Deepstream-style Bounding box.

height: float#
left: float#
top: float#
width: float#
pythia.utils.maskrcnn._gen_clips(width: ndarray, original_width: int, height: ndarray, original_height: int) → Tuple[ndarray, ndarray, ndarray, ndarray][source]#
pythia.utils.maskrcnn._gen_idxs(original_width: int, left: ndarray, right: ndarray, top: ndarray, bottom: ndarray) → Tuple[ndarray, ndarray, ndarray, ndarray][source]#
pythia.utils.maskrcnn._gen_ranges(original_height: int, original_width: int, target_height: int, target_width: int) → Tuple[ndarray, ndarray][source]#
pythia.utils.maskrcnn._interpolate(width: ndarray, left: ndarray, height: ndarray, top: ndarray, left_top_val: ndarray, right_top_val: ndarray, left_bottom_val: ndarray, right_bottom_val: ndarray) → ndarray[source]#
pythia.utils.maskrcnn._take_vals(src, *idxmats)[source]#
pythia.utils.maskrcnn.extract_maskrcnn_mask(obj_meta: NvDsObjectMeta) → ndarray[source]#

Extract maskrcnn mask from deepstream object.

Parameters:

obj_meta – Deepstream object meta from detection.

Returns:

A 2d binary mask of np.uint8 valued 0 and 255.

Example

>>> obj_meta = pyds.NvDsObjectMeta.cast(...)
>>> mask = extract_maskrcnn_mask(obj_meta)
>>> mask.shape, mask.dtype
((300,100), dtype('uint8'))

See also

resize_mask_vec for the internal implementation.

pythia.utils.maskrcnn.polygon_to_bbox(polygons: List[List[int]], top: float, left: float) → DsBBox[source]#

Generate bounding box from polygons.

The polygons are assumed to be relative to a mask.

Parameters:

  • polygons – collection of polygons, where each polygon is a sequence of the form ‘y0,x0,y1,x1,…,yn,xn’.

  • top – top offset from the mask.

  • left – left offset from the mask.

Returns:

Bounding box which circumscribes the mask.

pythia.utils.maskrcnn.resize_mask_vec(src: ndarray, src_shape: Tuple[int, int], target_shape: Tuple[int, int], threshold: float) → ndarray[source]#

Resize mask from original deepstream object into numpy array.

Parameters:

  • src – Mask array from deepstream object.

  • src_shape – Shape of the original mask in (height,width) format.

  • target_shape – Shape of the target mask in (height,width) format.

  • threshold – Threshold for the mask.

Returns:

A 2d binary mask of np.uint8 valued 0 and 255.

See also

  • extract_maskrcnn_mask in this module for sample usage from

    deepstream.

  • resizeMask function at

    sample_apps/deepstream-mrcnn-app/deepstream_mrcnn_test.cpp