fastmask-seg

Unique colors, pixel counts, per-color boolean masks, and bounding boxes from segmentation masks stored as images.

import cv2
import numpy as np
import fastmask_seg as fms

image = cv2.imread("instance_mask.png")  # (H, W, 3) uint8, one solid color per instance

result = fms.unique_color_masks(image)
result.colors        # (K, 3) uint8 colors present, in BGR order because OpenCV loads BGR
result.color_counts  # (K,) int64 pixel count per color
result.masks         # (K, H, W) bool, one mask per color

road = fms.get_color_mask(image, color=np.array([128, 64, 128], np.uint8))  # (H, W) bool
boxes = fms.masks_to_bboxes(result.masks)  # (K, 4) int64 rows of (rmin, rmax, cmin, cmax)

The library targets instance and semantic segmentation masks, where each class or instance is one solid color. Its fast paths rely on what those images look like: few distinct colors in long runs of identical pixels.

Install

pip install fastmask-seg            # NumPy only
pip install "fastmask-seg[numba]"   # adds compiled kernels for 3-channel images

Requires Python 3.12+ and NumPy 2.0+. The numba extra gives the same results as NumPy, with color analysis 1.3-2x faster and masks of named colors 2-4x faster (details).

API

fm = fms.FastMask(image)             # computes on first access and caches, so later queries reuse the analysis
fm.unique_colors                     # (K, C) uint8 colors in the image's channel layout, ordered by color ID
fm.unique_color_ids                  # (K,) sorted color IDs
fm.unique_color_counts               # (K,) int64 pixel counts
fm.label_image                       # (H, W) index of each pixel's color in unique_color_ids
fm.color_id_image                    # (H, W) color ID of each pixel
fm.get_color_masks(colors=palette)   # ColorMasks(masks, num_masks), one mask per requested color
fm.get_color_masks()                 # every color in the image
fm.get_color_mask(color=c)           # one (H, W) bool mask
fm.iter_color_masks()                # one (H, W) mask at a time, for many colors on large images

# Each FastMask member is also a function that takes an image or a FastMask:
fms.unique_colors(image)
fms.get_color_masks(image, colors=palette)
fms.unique_color_masks(image)        # UniqueColorMasks(colors, color_ids, color_counts, masks, num_masks)

fms.masks_to_bboxes(masks)           # (K, 4) boxes of a (K, H, W) bool stack
fms.mask_to_bbox(mask)               # (rmin, rmax, cmin, cmax) of one mask

fms.colors_to_color_ids(colors)      # (..., 3) or (..., 4) uint8 colors to uint32 IDs
fms.color_ids_to_colors(ids, fms.ColorFormat.COLOR)
fms.sort_colors(colors)              # colors in color ID order

• Images are uint8 arrays of shape (H, W), (H, W, 3), or (H, W, 4), as cv2.imread or np.asarray(PIL.Image.open(...)) give them. In a grayscale image the gray value is the color.
• Colors follow the image's channel order: BGR for OpenCV, RGB for Pillow. A color ID packs a color into one uint32 with channel 0 in the lowest byte.
• Mask requests take colors= or color_ids=, and with neither they cover every color in the image. A color absent from the image gives an all-False mask.
• Bounding boxes have exclusive ends, so mask[rmin:rmax, cmin:cmax] crops a mask to its pixels. An empty mask raises ValueError.
• Functions and FastMask: a function given an image analyzes it from scratch on each call. Wrap the image in a FastMask to ask several questions about one image, and pass that object to the functions when you prefer their form.
• Arguments are used as given. The docstrings state each required shape and dtype, and other inputs surface as NumPy errors or wrong results.

Processing many images

Allocate the mask buffer once and pass it as out=, so each image writes into memory the process already holds:

masks_buf = np.empty((64, H, W), dtype=bool)  # room for up to 64 colors per image

for path in paths:
    result = fms.unique_color_masks(cv2.imread(path), out=masks_buf)
    ...  # result.masks is masks_buf[:result.num_masks], overwritten by the next image

For a fixed class palette, pass the palette straight to get_color_masks(image, colors=palette, out=...); with the numba extra this is the fastest path. docs/performance.md covers the backends, the NumPy-path color_id_buffer, caching, and benchmarks.

Development

make dev     # uv sync --dev
make check   # ruff, isort, mypy --strict
make format  # isort + ruff format

License

MIT