TILEs

Python toolkit for decomposing images into collections of colored polygonal tiles. It provides several partitioning strategies, including quadtrees, Voronoi trees, recursive slicing, and random polygons. Tile boundaries can be guided by local image statistics, a heightmap, or custom stopping and color policies.

The resulting tiles can be rendered as images or plotted as vector-like polygon collections.

Quick start

from tiles import load_images, save_tileset
from tiles.partitioners.quadtree import QuadtreePartitioner
from tiles.recursion import grow_tree
from tiles.stopping_policies import UniformColorStop
from tiles.color_policies import MeanRGBColor

context = load_images("input.jpg")
tile_set = grow_tree(
    context,
    partitioner=QuadtreePartitioner(),
    stop_policy=UniformColorStop(standard_deviation=30),
    color_policy=MeanRGBColor(),
    seed=42,
)
save_tileset(tile_set, "tiles.png", width=8)

The same TileSet can be rendered as a Matplotlib figure with tileset_to_figure, or saved directly with save_tileset.

Color policies

The color_policy argument controls the representative color assigned to each tile. Without one, TILEs uses the region's mean RGB color.

Stop policies

The stop_policy determines when a region should stop being subdivided. For example, MaxLevelStop limits recursion depth, while UniformColorStop stops when a region's color variation is below a threshold. Other built-in policies can use heightmap contrast, heightmap range, target tile size, perceptual uniformity, or stochastic stopping. Policies can also be combined with AllOf and AnyOf.

from tiles.stopping_policies import AnyOf, MaxLevelStop, UniformColorStop

stop_policy = AnyOf(
    UniformColorStop(standard_deviation=30),
    MaxLevelStop(max_level=6),
)

Examples

Model: @maleficent_model

Partitioning strategies

Color policies

Performance notes

• Quadtree and recursive slicing compute region statistics in constant time from summed-area tables (tiles.integral). The tables cost about 40 bytes per pixel and are skipped above tiles.integral.INTEGRAL_MAX_PIXELS (4 million pixels); results are identical either way, only slower.
• Each region's visible pixels and polygon mask are computed once and shared by the stop policy, the colour policy and the tile image.
• Tile images are rendered when accessed, not stored. Tile.image re-renders on every access, which keeps large tile sets light in memory.

See a Notebook for more information.

This was originally written to use with PILes