Run-Length Encoded Mask Operations

This library is an extended and improved version of the COCO API's pycocotools.mask module (which was originally written by Piotr Dollár and Tsung-Yi Lin).

It offers the following additional features:

• Further set operations (complement, difference, symmetric difference) in RLE, without decoding
• Mask cropping and padding in RLE
• Connected components extraction in RLE (rlemasklib.connected_components)
• Direct creation of full and empty masks in RLE
• Decompression of COCO's compressed RLE format to integer run-lengths, and vice versa
• Extra compression (optional) using gzip on top of the LEB128-like encoding used by the COCO API (~40% reduction beyond the COCO compression)
• More streamlined API

List of functions

Encoding / decoding (between run lengths and binary masks)

• rle_mask = rlemasklib.encode(binary_mask): Encode a binary mask as RLE
• binary_mask = rlemasklib.decode(rle_mask): Decode an RLE mask to a binary mask

Compression / decompression (between compressed and uncompressed run lengths)

• rle_mask = rlemasklib.compress(rle_mask): Compress an RLE mask using LEB128 (and optionally gzip)
• rle_mask = rlemasklib.decompress(rle_mask): Decompress an RLE mask from LEB128 or gzip to an array of integers (run-lengths)

Initialization

• rle_mask = rlemasklib.empty(imshape): Create an empty RLE mask of given size
• rle_mask = rlemasklib.full(imshape): Create a full RLE mask of given size

Set operations

• rle_mask = rlemasklib.intersection(rle_masks): Compute the intersection of multiple RLE masks.
• rle_mask = rlemasklib.union(rle_masks): Compute the union of multiple RLE masks.
• rle_mask = rlemasklib.complement(rle_mask): Compute the complement of an RLE mask.
• rle_mask = rlemasklib.difference(rle_mask1, rle_mask2): Compute the difference of two RLE masks.
• rle_mask = rlemasklib.symmetric_difference(rle_mask1, rle_mask2): Compute the symmetric difference of two RLE masks.

Measurements

• area = rlemasklib.area(rle_mask): Compute the area of an RLE mask
• centroid = rlemasklib.centroid(rle_mask): Compute the centroid of an RLE mask (or multiple masks). Returns [x, y] coordinates. The centroid is the average position of the foreground pixels.
• iou = rlemasklib.iou(rle_masks): Compute the intersection-over-union of multiple (typically two) RLE masks.

Crop / pad / shift by offset

• rle_mask = rlemasklib.crop(rle_mask, bbox): Crop an RLE mask to a given bounding box, yielding a mask with smaller height and/or width.
• rle_mask = rlemasklib.pad(rle_mask, paddings, value=0): Pad an RLE mask with given amount of [left, right, top, bottom] pixels with given value (0 or 1).
• rle_mask = rlemasklib.shift(rle_mask, offset, border_value=0): Shift an RLE mask by a given pixel offset [dx, dy], filling the border with a given value.

Connected components

• rle_masks = rlemasklib.connected_components(rle_mask, connectivity=4, min_size=1): Extract the connected components of the foreground from an RLE mask. Connectivity can be 4 or 8. Minimum size can be set to filter out small components.
• rle_mask = rlemasklib.largest_connected_component(rle_mask, connectivity=4): Returns the largest connected component of the foreground from an RLE mask. Returns None if there is no foreground.
• rle_mask = rlemasklib.remove_small_components(rle_mask, connectivity, min_size): Remove small connected components from the foreground of an RLE mask.
• rle_mask = rlemasklib.fill_small_holes(rle_mask, connectivity, min_size): Fill small holes (connected components of the background) in an RLE mask.

Conversions (bounding box, polygon)

• [x_start, y_start, width, height] = rlemasklib.to_bbox(rle_mask): Convert an RLE mask to a bounding box.
• rle_mask = rlemasklib.from_bbox([x_start, y_start, width, height], imshape): Convert a bounding box to an RLE mask inside a given image size.
• rle_mask = rlemasklib.from_polygon(polygon, imshape): Convert a polygon to an RLE mask inside a given image size.