Is there a way in matplotlib to check which artists are in the currently displayed area of the axes?

Question

I have a program with an interactive figure where occasionally many artists are drawn. In this figure, you can also zoom and pan using the mouse. However, the performace during zooming an panning is not very good because every artist is always redrawn. Is there a way to check which artists are in the currently displayed area and only redraw those? (In the example below the perfomace is still relatively good, but it can be made arbitrarily worse by using more or more complex artists)

I had a similar performace problem with the hover method that whenever it was called it ran canvas.draw() at the end. But as you can see I found a neat workaround for that by making use of caching and restoring the background of the axes (based on this). This significantly improved the performace and now even with many artists it runs very smooth. Maybe there is a similar way of doing this but for the pan and zoom method?

Sorry for the long code sample, most of it is not directly relevant for the question but necessary for a working example to highlight the issue.

EDIT

I updated the MWE to something that is more representative of my actual code.

import numpy as np
import sys
import matplotlib.pyplot as plt
from matplotlib.backends.backend_qt5agg import \
    FigureCanvasQTAgg
import matplotlib.patheffects as PathEffects
from matplotlib.text import Annotation
from matplotlib.collections import LineCollection

from PyQt5.QtWidgets import QApplication, QVBoxLayout, QDialog


def check_limits(base_xlim, base_ylim, new_xlim, new_ylim):
    if new_xlim[0] < base_xlim[0]:
        overlap = base_xlim[0] - new_xlim[0]
        new_xlim[0] = base_xlim[0]
        if new_xlim[1] + overlap > base_xlim[1]:
            new_xlim[1] = base_xlim[1]
        else:
            new_xlim[1] += overlap
    if new_xlim[1] > base_xlim[1]:
        overlap = new_xlim[1] - base_xlim[1]
        new_xlim[1] = base_xlim[1]
        if new_xlim[0] - overlap < base_xlim[0]:
            new_xlim[0] = base_xlim[0]
        else:
            new_xlim[0] -= overlap
    if new_ylim[1] < base_ylim[1]:
        overlap = base_ylim[1] - new_ylim[1]
        new_ylim[1] = base_ylim[1]
        if new_ylim[0] + overlap > base_ylim[0]:
            new_ylim[0] = base_ylim[0]
        else:
            new_ylim[0] += overlap
    if new_ylim[0] > base_ylim[0]:
        overlap = new_ylim[0] - base_ylim[0]
        new_ylim[0] = base_ylim[0]
        if new_ylim[1] - overlap < base_ylim[1]:
            new_ylim[1] = base_ylim[1]
        else:
            new_ylim[1] -= overlap

    return new_xlim, new_ylim


class FigureCanvas(FigureCanvasQTAgg):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.bg_cache = None

    def draw(self):
        ax = self.figure.axes[0]
        hid_annotation = False
        if ax.annot.get_visible():
            ax.annot.set_visible(False)
            hid_annotation = True
        hid_highlight = False
        if ax.last_artist:
            ax.last_artist.set_path_effects([PathEffects.Normal()])
            hid_highlight = True
        super().draw()
        self.bg_cache = self.copy_from_bbox(self.figure.bbox)
        if hid_highlight:
            ax.last_artist.set_path_effects(
                [PathEffects.withStroke(
                    linewidth=7, foreground="c", alpha=0.4
                )]
            )
            ax.draw_artist(ax.last_artist)
        if hid_annotation:
            ax.annot.set_visible(True)
            ax.draw_artist(ax.annot)

        if hid_highlight:
            self.update()


def position(t_, coeff, var=0.1):
    x_ = np.random.normal(np.polyval(coeff[:, 0], t_), var)
    y_ = np.random.normal(np.polyval(coeff[:, 1], t_), var)

    return x_, y_


class Data:
    def __init__(self, times):
        self.length = np.random.randint(1, 20)
        self.t = np.sort(
            np.random.choice(times, size=self.length, replace=False)
        )
        self.vel = [np.random.uniform(-2, 2), np.random.uniform(-2, 2)]
        self.accel = [np.random.uniform(-0.01, 0.01), np.random.uniform(-0.01,
                                                                      0.01)]
        x0, y0 = np.random.uniform(0, 1000, 2)
        self.x, self.y = position(
            self.t, np.array([self.accel, self.vel, [x0, y0]])
        )


class Test(QDialog):
    def __init__(self):
        super().__init__()
        self.fig, self.ax = plt.subplots()
        self.canvas = FigureCanvas(self.fig)
        self.artists = []
        self.zoom_factor = 1.5
        self.x_press = None
        self.y_press = None
        self.annot = Annotation(
            "", xy=(0, 0), xytext=(-20, 20), textcoords="offset points",
            bbox=dict(boxstyle="round", fc="w", alpha=0.7), color='black',
            arrowprops=dict(arrowstyle="->"), zorder=6, visible=False,
            annotation_clip=False, in_layout=False,
        )
        self.annot.set_clip_on(False)
        setattr(self.ax, 'annot', self.annot)
        self.ax.add_artist(self.annot)
        self.last_artist = None
        setattr(self.ax, 'last_artist', self.last_artist)

        self.image = np.random.uniform(0, 100, 1000000).reshape((1000, 1000))
        self.ax.imshow(self.image, cmap='gray', interpolation='nearest')
        self.times = np.linspace(0, 20)
        for i in range(1000):
            data = Data(self.times)
            points = np.array([data.x, data.y]).T.reshape(-1, 1, 2)
            segments = np.concatenate([points[:-1], points[1:]], axis=1)
            z = np.linspace(0, 1, data.length)
            norm = plt.Normalize(z.min(), z.max())
            lc = LineCollection(
                segments, cmap='autumn', norm=norm, alpha=1,
                linewidths=2, picker=8, capstyle='round',
                joinstyle='round'
            )
            setattr(lc, 'data_id', i)
            lc.set_array(z)
            self.ax.add_artist(lc)
            self.artists.append(lc)
        self.default_xlim = self.ax.get_xlim()
        self.default_ylim = self.ax.get_ylim()

        self.canvas.draw()

        self.cid_motion = self.fig.canvas.mpl_connect(
            'motion_notify_event', self.motion_event
        )
        self.cid_button = self.fig.canvas.mpl_connect(
            'button_press_event', self.pan_press
        )
        self.cid_zoom = self.fig.canvas.mpl_connect(
            'scroll_event', self.zoom
        )

        layout = QVBoxLayout()
        layout.addWidget(self.canvas)
        self.setLayout(layout)

    def zoom(self, event):
        if event.inaxes == self.ax:
            scale_factor = np.power(self.zoom_factor, -event.step)
            xdata = event.xdata
            ydata = event.ydata
            cur_xlim = self.ax.get_xlim()
            cur_ylim = self.ax.get_ylim()
            x_left = xdata - cur_xlim[0]
            x_right = cur_xlim[1] - xdata
            y_top = ydata - cur_ylim[0]
            y_bottom = cur_ylim[1] - ydata

            new_xlim = [
                xdata - x_left * scale_factor, xdata + x_right * scale_factor
            ]
            new_ylim = [
                ydata - y_top * scale_factor, ydata + y_bottom * scale_factor
            ]
            # intercept new plot parameters if they are out of bounds
            new_xlim, new_ylim = check_limits(
                self.default_xlim, self.default_ylim, new_xlim, new_ylim
            )

            if cur_xlim != tuple(new_xlim) or cur_ylim != tuple(new_ylim):
                self.ax.set_xlim(new_xlim)
                self.ax.set_ylim(new_ylim)

                self.canvas.draw_idle()

    def motion_event(self, event):
        if event.button == 1:
            self.pan_move(event)
        else:
            self.hover(event)

    def pan_press(self, event):
        if event.inaxes == self.ax:
            self.x_press = event.xdata
            self.y_press = event.ydata

    def pan_move(self, event):
        if event.inaxes == self.ax:
            xdata = event.xdata
            ydata = event.ydata
            cur_xlim = self.ax.get_xlim()
            cur_ylim = self.ax.get_ylim()
            dx = xdata - self.x_press
            dy = ydata - self.y_press
            new_xlim = [cur_xlim[0] - dx, cur_xlim[1] - dx]
            new_ylim = [cur_ylim[0] - dy, cur_ylim[1] - dy]

            # intercept new plot parameters that are out of bound
            new_xlim, new_ylim = check_limits(
                self.default_xlim, self.default_ylim, new_xlim, new_ylim
            )

            if cur_xlim != tuple(new_xlim) or cur_ylim != tuple(new_ylim):
                self.ax.set_xlim(new_xlim)
                self.ax.set_ylim(new_ylim)

                self.canvas.draw_idle()

    def update_annot(self, event, artist):
        self.ax.annot.xy = (event.xdata, event.ydata)
        text = f'Data #{artist.data_id}'
        self.ax.annot.set_text(text)
        self.ax.annot.set_visible(True)
        self.ax.draw_artist(self.ax.annot)

    def hover(self, event):
        vis = self.ax.annot.get_visible()
        if event.inaxes == self.ax:
            ind = 0
            cont = None
            while (
                ind in range(len(self.artists))
                and not cont
            ):
                artist = self.artists[ind]
                cont, _ = artist.contains(event)
                if cont and artist is not self.ax.last_artist:
                    if self.ax.last_artist is not None:
                        self.canvas.restore_region(self.canvas.bg_cache)
                        self.ax.last_artist.set_path_effects(
                            [PathEffects.Normal()]
                        )
                        self.ax.last_artist = None
                    artist.set_path_effects(
                        [PathEffects.withStroke(
                            linewidth=7, foreground="c", alpha=0.4
                        )]
                    )
                    self.ax.last_artist = artist
                    self.ax.draw_artist(self.ax.last_artist)
                    self.update_annot(event, self.ax.last_artist)
                ind += 1

            if vis and not cont and self.ax.last_artist:
                self.canvas.restore_region(self.canvas.bg_cache)
                self.ax.last_artist.set_path_effects([PathEffects.Normal()])
                self.ax.last_artist = None
                self.ax.annot.set_visible(False)
        elif vis:
            self.canvas.restore_region(self.canvas.bg_cache)
            self.ax.last_artist.set_path_effects([PathEffects.Normal()])
            self.ax.last_artist = None
            self.ax.annot.set_visible(False)
        self.canvas.update()
        self.canvas.flush_events()


if __name__ == '__main__':
    app = QApplication(sys.argv)
    test = Test()
    test.show()
    sys.exit(app.exec_())

I don't understand the problem. Since artists that are outside the axes are not drawn anyways, they will also not slow down anything. — ImportanceOfBeingErnest, Feb 27 '20 at 11:22
So you are saying there already is a routine that checks which of the artists can be seen so only the visible ones are actually drawn? Maybe this routine is what is computationally very expensive? Because you can easily see a difference in performance if you try the following e.g.: with my 1000 artist WME above, zoom in on a single artist and pan around. You will notice a significant delay. Now do the same but plot only 1 (or even 100) artist(s) and you will see that there is almost no delay. — mapf, Feb 27 '20 at 12:21
Well, the question is, are you able to write a more efficient routine? In a simple case, maybe. So you can check which artists are within the view limits and set all other invisible. If the check just compares the center coordinates of the dots, that's faster. But that would make you loose the dot if only its center is outside but slightly less than half of it would still be inside the view. That being said, the main problem here is that there are 1000 artists in the axes. If instead, you used only one single `plot` with all points, the problem would not occur. — ImportanceOfBeingErnest, Feb 27 '20 at 13:25
Yeah absolutely true. It's just that my premise was wrong. I thought the reason for the bad performace was that all artists are aways drawn independent of whether they can be seen or not. Thus I thought a smart routine that only draws the artists that are going to be seen would improve the performance but apparently such routine is already in place, so I guess there is not much that can be done here. I'm pretty sure I won't be able to write a more efficient routine, at least for a general case. — mapf, Feb 27 '20 at 13:45
In my case however, I am actually dealing with linecollections (plus an image in the background) and as you already said, even if it were just dots as in my MWE, simply checking if the coordinates are inside the axes is not enough. Maybe I should update the MWE accordingly to make it clearer. — mapf, Feb 27 '20 at 13:49

score 1 · Answer 1 · answered Feb 27 '20 at 10:39

1

You can find which artists are in the current area of the axes if you focus on the data the artists are plotting.

For example if you put your points data (a and b arrays) in a numpy array like this:

self.points = np.random.randint(0, 100, (1000, 2))

you can get the list of points inside the current x and y limits:

xmin, xmax = self.ax.get_xlim()
ymin, ymax = self.ax.get_ylim()

p = self.points

indices_of_visible_points = (np.argwhere((p[:, 0] > xmin) & (p[:, 0] < xmax) & (p[:, 1] > ymin) &  (p[:, 1] < ymax))).flatten()

you can use indices_of_visible_points to index your related self.artists list

answered Feb 27 '20 at 10:39

Guglie

2,121
1
24
43

Thank you for your answer! Unfortunately, this only works in case the artists are single points. It already doesn't work anymore if the artists are lines. E.g. image a line defined by only two points where the points lie outside of the axes limits, however the line connecting the points is intersecting the axes frame. Maybe I should edit the MWE accordingly so it more obvious. – mapf Feb 27 '20 at 11:03
To me the approach is the same, **focus on the data**. If the artists are lines you can additionally check for intersection with the view rectangle. If you're plotting curves, probably you're sampling them at fixed intervals reducing them to line segments. By the way, can you give a more realistic sample of what are you plotting? – Guglie Feb 27 '20 at 12:46

Is there a way in matplotlib to check which artists are in the currently displayed area of the axes?

1 Answers1