geowatch.tasks.tracking.polygon_extraction module

class geowatch.tasks.tracking.polygon_extraction.PolygonExtractor(heatmap_thwc, heatmap_time_intervals=None, bounds=None, classes=None, config=None)

Bases: object

Given a timesequence of heatmaps, extract spatially static polygons.

This class is being developed on dev/refine-ac-polys in the file: geowatch/tasks/tracking/polygon_extraction.py

Distributed Tweaking on:

https://colab.research.google.com/drive/1NEJpm36LviesZb45qy59myezi7JHu0bI#scrollTo=G8kHgCXSI3VS

Example

>>> from geowatch.tasks.tracking.polygon_extraction import *  # NOQA
>>> cls = PolygonExtractor
>>> self = PolygonExtractor.demo(real_categories=True)
>>> self.config['algo'] = 'crall'
>>> print(f'self.heatmap_thwc.shape={self.heatmap_thwc.shape}')

>>> polys, info = self.predict_polygons(return_info=True)
>>> label_mask = info['label_mask']

>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> from geowatch.utils import util_kwimage
>>> kwplot.autompl()
>>> stacked = self.draw_timesequence()
>>> canvas = label_mask.colorize()
>>> for p in polys:
>>>     canvas = p.draw_on(canvas, fill=0, edgecolor='black')
>>> pnum_ = kwplot.PlotNums(nSubplots=2)
>>> kwplot.imshow(stacked, pnum=pnum_(), title='Colorized Time Series')
>>> kwplot.imshow(canvas, pnum=pnum_(), title='Cluster Labels')
>>> kwplot.show_if_requested()

Parameters:

• heatmap_thwc (ndarray) – An heatmap with dimensions [T, H, W, C] where T is number of timesteps and C are feature map reponces for each channel.

• heatmap_time_intervals (None | List[Tuple[datetime, datetime]]) – A list of length T rows. Each row corresponds to a timestep in the heatmaps, and contains a tuple of two datetimes: corresponding to the start and end of times that information corresponds to. If a heatmap only corrseponds to a single time, then start and end will be the same.

• bounds (kwimage.MultiPolygon | None) – if given, this is the polygon detected at BAS time.

• classes (None | List[str]) – If specified, this corresponds to the class in each position of the C dimension.

• config (None | Dict) – optional configuration

predict()

Predict the spatial polygons

Returns:

A single [H, W] integer map indicating a cluster label for each pixel. I.e. a spatial segmentation of the sites.

Return type:

ndarray

predict_polygons(return_info=False)

show()

draw_intermediate()

draw_timesequence()

classmethod demo(**kwargs)

Create an instance of the problem on toy data.

Parameters:

**kwargs – passed to PolygonExtractor.demodata()

Returns:

PolygonExtractor

classmethod demodata(real_categories=False, rng=0)

Create toydata to test and demo the API

Parameters:

• real_categories (bool) – if False, use fake cateogires

• rng (int | None) – random seed, or None to use global seed.

Returns:

A dictionary that can be used as kwargs

to construct and instance of this class.

Return type:

Dict[str, Any]

geowatch.tasks.tracking.polygon_extraction.impute_nans(data)

interpolate to fill nan values

TODO: tests

data = np.random.rand(5, 3, 3, 2) data[data < 0.1] = np.nan

data = np.random.rand(54, 295, 296, 5) data[data < 0.1] = np.nan

import timerit ti = timerit.Timerit(1, bestof=1, verbose=3)

for timer in ti.reset(‘impute_nans2’):

with timer:

impute_nans2(data)

for timer in ti.reset(‘impute_nans’):

with timer:

impute_nans(data)

geowatch.tasks.tracking.polygon_extraction.impute_nans2(data)

class geowatch.tasks.tracking.polygon_extraction.FeatureCube(heatmap_thwc, heatmap_time_intervals=None, classes=None)

Bases: NiceRepr

Container for a [T, H, W, C] heatmap and corresponding T-lengthed time intervals.

take_channels(channels)

window_max(window)

classmethod demo()

draw()

class geowatch.tasks.tracking.polygon_extraction.TimeIntervalSequence(iterable=(), /)

Bases: list

A list of non-overlapping time intervals

apply_grouping(groupxs)

classmethod coerce(data)

class geowatch.tasks.tracking.polygon_extraction.TimeInterval(*intervals)

Bases: Interval

Represents an interval in time.

Example

from geowatch.tasks.tracking.polygon_extraction import * # NOQA a = TimeInterval.coerce((‘2020-01-01’, ‘2020-02-01’)) b = TimeInterval.coerce((‘2020-02-01’, ‘2020-03-01’)) c = TimeInterval.coerce((‘2020-03-01’, ‘2020-03-15’)) d = TimeInterval.coerce((‘2020-03-07’, ‘2020-04-01’)) e = TimeInterval.coerce((‘2020-04-15’, ‘2020-05-01’))

Create a disjunction of zero, one or more intervals.

Parameters:

intervals – zero, one or more intervals.

property start

property stop

classmethod coerce(data)

classmethod closed(start, stop=None)

classmethod random()

geowatch.tasks.tracking.polygon_extraction.toydata_demo()

Example

from geowatch.tasks.tracking.polygon_extraction import * # NOQA cls = PolygonExtractor(heatmaps, bounds) self = PolygonExtractor.demo()

geowatch.tasks.tracking.polygon_extraction.real_data_demo_case_1()

geowatch.tasks.tracking.polygon_extraction.real_data_demo_case_2()

geowatch.tasks.tracking.polygon_extraction.real_data_demo_case_3()

geowatch.tasks.tracking.polygon_extraction.real_data_demo_case_1_fixed()

geowatch.tasks.tracking.polygon_extraction.generate_real_example()

class geowatch.tasks.tracking.polygon_extraction.LabelMask(data)

Bases: object

colorize()

resize(scale=None, dsize=None)

classmethod demo()

to_multi_polygons()

geowatch.tasks.tracking.polygon_extraction.coco_make_track_gdf(coco_dset, video_id, resolution=None)

geowatch.tasks.tracking.polygon_extraction.imread_many(dset, gids, channels=None, space='video', resolution=None, workers=0)