cvdatasets/nabirds/dataset/part.py

import numpy as np
from matplotlib import pyplot as plt

from contextlib import contextmanager
from matplotlib.patches import Rectangle
from abc import ABC, abstractmethod, abstractproperty

from . import utils

class BaseParts(ABC):

	def __getitem__(self, i):
		return self._parts[i]

	@property
	def selected(self):
		return np.array([p.is_visible for p in self._parts], dtype=bool)

	@property
	def selected_idxs(self):
		return np.where(self.selected)[0]

	def select(self, idxs):
		if isinstance(idxs, np.ndarray) and idxs.dtype == bool:
			# a mask is present, so convert it to indeces
			idxs = np.where(idxs)[0]

		for p in self._parts:
			p.is_visible = p._id in idxs


	def invert_selection(self):
		self.select(np.logical_not(self.selected))

	def offset(self, dx, dy):
		for p in self._parts:
			p.x += dx
			p.y += dy

	def visible_locs(self):
		vis = [(p._id, p.xy) for p in self._parts if p.is_visible]
		idxs, xy = zip(*vis)
		return np.array(idxs), np.array(xy).T

	def visible_crops(self, *args, **kwargs):
		return np.array([p.crop(*args, **kwargs) for p in self._parts])

	def plot(self, cmap=plt.cm.jet, **kwargs):
		for i, p in enumerate(self._parts):
			p.plot(color=cmap(i/len(self._parts)), **kwargs)

	def reveal(self, im, ratio, *args, **kwargs):
		res = np.zeros_like(im)

		for part in self._parts:
			if not part.is_visible: continue
			x, y, crop = part.reveal(im, ratio=ratio, *args, **kwargs)
			h, w, _ = crop.shape
			res[y:y+h, x:x+w] = crop

		return res

class Parts(BaseParts):

	def __init__(self, image, part_annotations, rescale_size):
		super(Parts, self).__init__()
		self._parts = [BasePart.new(image, a, rescale_size) for a in part_annotations]


class UniformParts(BaseParts):

	def __init__(self, image, ratio):
		super(UniformParts, self).__init__()
		self._parts = list(self.generate_parts(image, ratio))

	def generate_parts(self, im, ratio, round_op=np.floor):
		h, w, c = utils.dimensions(im)

		part_w = round_op(w * ratio).astype(np.int32)
		part_h = round_op(h * ratio).astype(np.int32)

		n, m = w // part_w, h // part_h

		# fit best possible part_w and part_h
		part_w = int(w / n)
		part_h = int(h / m)

		for i in range(n*m):
			row, col = np.unravel_index(i, (n, m))
			x, y = col * part_w, row * part_h

			yield BBoxPart(im, [i, x, y, part_w, part_h])

class BasePart(ABC):

	@staticmethod
	def new(image, annotation, rescale_size=-1):
		if len(annotation) == 4:
			return LocationPart(image, annotation, rescale_size)
		elif len(annotation) == 5:
			return BBoxPart(image, annotation, rescale_size)
		else:
			raise ValueError("Unknown part annotation format: {}".format(annotation))

	def rescale(self, image, annotation, rescale_size):
		if rescale_size is not None and rescale_size > 0:
			h, w, c = utils.dimensions(image)
			scale = np.array([w, h]) / rescale_size
			xy = annotation[1:3]
			xy = xy * scale
			annotation[1:3] = xy

		return annotation

	@property
	def is_visible(self):
		return self._is_visible

	@is_visible.setter
	def is_visible(self, value):
		self._is_visible = bool(value)

	@property
	def xy(self):
		return np.array([self.x, self.y])

	def crop(self, im, w, h, padding_mode="edge", is_location=True):
		if not self.is_visible:
			_, _, c = utils.dimensions(im)
			return np.zeros((h, w, c), dtype=np.uint8)

		x, y = self.xy
		pad_h, pad_w = h // 2, w // 2

		padded_im = np.pad(im, [(pad_h, pad_h), (pad_w, pad_w), [0,0]], mode=padding_mode)
		x0, y0 = x + pad_w, y + pad_h

		if is_location:
			x0, y0 = x0 - w // 2, y0 - h // 2

		return padded_im[y0:y0+h, x0:x0+w]


	def plot(self, **kwargs):
		return

class LocationPart(BasePart):
	DEFAULT_RATIO = np.sqrt(49 / 400) # 0.35

	def __init__(self, image, annotation, rescale_size=None):
		super(LocationPart, self).__init__()

		annotation = self.rescale(image, annotation, rescale_size)
		# here x,y are the center of the part
		self._id, self.x, self.y, self.is_visible = annotation
		self._ratio = LocationPart.DEFAULT_RATIO

	def crop(self, image, ratio=None, padding_mode="edge", *args, **kwargs):
		ratio = ratio or self._ratio
		_h, _w, c = utils.dimensions(image)
		w, h = int(_w * ratio), int(_h * ratio)
		return super(LocationPart, self).crop(image, w, h,
				padding_mode, is_location=True)

	def reveal(self, im, ratio=None, *args, **kwargs):
		_h, _w, c = utils.dimensions(im)
		w, h = int(_w * ratio), int(_h * ratio)
		x,y = self.xy
		x, y = max(x - w // 2, 0), max(y - h // 2, 0)
		return x, y, im[y:y+h, x:x+w]


	def plot(self, im, ax, ratio, fill=False, linestyle="--", **kwargs):
		if not self.is_visible: return
		x, y = self.xy
		_h, _w, c = utils.dimensions(im)
		w, h = int(_w * ratio), int(_h * ratio)
		ax.add_patch(Rectangle(
			(x-w//2, y-h//2), w, h,
			fill=fill, linestyle=linestyle,
			**kwargs
		))

class BBoxPart(BasePart):

	def __init__(self, image, annotation, rescale_size=None):
		super(BBoxPart, self).__init__()

		annotation = self.rescale(image, annotation, rescale_size)
		# here x,y are top left corner of the part
		self._id, self.x, self.y, self.w, self.h = annotation
		self.is_visible = True

	def rescale(self, image, annotation, rescale_size):
		if rescale_size is not None and rescale_size > 0:
			annotation = super(BBoxPart, self).rescale(image, annotation, rescale_size)
			h, w, c = utils.dimensions(image)
			scale = np.array([w, h]) / rescale_size
			wh = annotation[3:5]
			wh = wh * scale
			annotation[3:5] = wh

		return annotation


	def crop(self, image, padding_mode="edge", *args, **kwargs):
		return super(BBoxPart, self).crop(image, self.w, self.h,
			padding_mode, is_location=False)

	def reveal(self, im, ratio, *args, **kwargs):
		_h, _w, c = utils.dimensions(im)
		x,y = self.xy
		return x, y, im[y:y+self.h, x:x+self.w]



	def plot(self, im, ax, ratio, fill=False, linestyle="--", **kwargs):
		ax.add_patch(Rectangle(
			(self.x, self.y), self.w, self.h,
			fill=fill, linestyle=linestyle,
			**kwargs
		))