|
|
|
|
|
|
|
|
|
import random
|
|
import torch
|
|
|
|
from .densepose_base import DensePoseBaseSampler
|
|
|
|
|
|
class DensePoseUniformSampler(DensePoseBaseSampler):
|
|
"""
|
|
Samples DensePose data from DensePose predictions.
|
|
Samples for each class are drawn uniformly over all pixels estimated
|
|
to belong to that class.
|
|
"""
|
|
|
|
def __init__(self, count_per_class: int = 8):
|
|
"""
|
|
Constructor
|
|
|
|
Args:
|
|
count_per_class (int): the sampler produces at most `count_per_class`
|
|
samples for each category
|
|
"""
|
|
super().__init__(count_per_class)
|
|
|
|
def _produce_index_sample(self, values: torch.Tensor, count: int):
|
|
"""
|
|
Produce a uniform sample of indices to select data
|
|
|
|
Args:
|
|
values (torch.Tensor): an array of size [n, k] that contains
|
|
estimated values (U, V, confidences);
|
|
n: number of channels (U, V, confidences)
|
|
k: number of points labeled with part_id
|
|
count (int): number of samples to produce, should be positive and <= k
|
|
|
|
Return:
|
|
list(int): indices of values (along axis 1) selected as a sample
|
|
"""
|
|
k = values.shape[1]
|
|
return random.sample(range(k), count)
|
|
|
