add scale_mask

[yuyu2172]

What this is

This code exploits sparsity of mask used for instance segmentation to speed up scale operation.
When scaling factor is an integer, there is no difference between resize.
When scaling factor is not an integer, there is one pixel error in horizontal or vertical direction.

Benchmark

The script below benchmarks performance in comparison to the normal resize.
The comparison is made by varying the number of instances in an image.

The figure shows the execution time in seconds.
Blue is the time taken by resize and orange is the execution time of scale_mask.

from chainercv.datasets import COCOInstanceSegmentationDataset
from chainercv import transforms
import PIL.Image
import numpy as np
import time
from chainercv.utils import mask_to_bbox
import matplotlib.pyplot as plt

from chainercv.utils import scale_mask


def old(mask, scale):
    _, H, W = mask.shape
    return transforms.resize(
        mask.astype(np.float32),
        (int(scale * H), int(scale * W)),
        interpolation=PIL.Image.NEAREST).astype(np.bool)


def measure_speed(func, args, n_call=20):
    exec_times = []
    for _ in range(n_call):
        start = time.time()
        func(*args)
        exec_times.append(time.time() - start)
    exec_times = np.sort(exec_times)
    return np.mean(exec_times[n_call // 10:-n_call // 10])


dataset = COCOInstanceSegmentationDataset(split='val', return_bbox=True)


count_to_id = {i: [] for i in range(100)}
for i in range(len(dataset)):
    label, = dataset.slice[:, ['label']][i]
    count_to_id[len(label)].append(i)

n = 40
old_times = np.zeros((n,))
new_times = np.zeros((n,))
for n_inst in range(1, n):
    if len(count_to_id[n_inst]) == 0:
        continue
    idx = count_to_id[n_inst][0]
    img = dataset[idx][0]
    mask = dataset[idx][1]
    bbox = mask_to_bbox(mask)
    assert len(bbox) == n_inst

    scale = 2
    H, W = mask.shape[1:]
    size = min(H, W) * scale

    old_times[n_inst] = measure_speed(old, (mask, scale))
    new_times[n_inst] = measure_speed(scale_mask, (mask, bbox, size))

    old_mask = old(mask, scale)
    fast_mask = scale_mask(mask, bbox, size)
    np.testing.assert_equal(fast_mask, old_mask)
    print('n_instance: {} old: {}  new: {}'.format(
        n_inst, old_times[n_inst], new_times[n_inst]))
plt.plot(old_times)
plt.plot(new_times)
plt.show()