Overview
Provides anti-aliased downsampling operations by combining blur (low-pass) filtering with pooling on 2D feature maps.
Description
This module is part of the Kornia library's filters subpackage. It implements blur pooling as described by Zhang (2019) for shift-invariant downsampling. The file contains three nn.Module classes (BlurPool2D, MaxBlurPool2D, EdgeAwareBlurPool2D) and their corresponding functional interfaces (blur_pool2d, max_blur_pool2d, edge_aware_blur_pool2d). BlurPool2D applies a Pascal (binomial) kernel followed by strided convolution for anti-aliased downsampling. MaxBlurPool2D combines max pooling with blur pooling. EdgeAwareBlurPool2D performs blur pooling while preserving edges using an edge-intensity threshold and dilation.
Usage
Import these classes or functions when you need to downsample feature maps or images in a shift-invariant manner, for example as a replacement for strided convolution or standard pooling in convolutional neural networks.
Code Reference
Source Location
Signature
class BlurPool2D(nn.Module):
def __init__(self, kernel_size: tuple[int, int] | int, stride: int = 2) -> None: ...
def forward(self, input: torch.Tensor) -> torch.Tensor: ...
class MaxBlurPool2D(nn.Module):
def __init__(self, kernel_size: tuple[int, int] | int, stride: int = 2,
max_pool_size: int = 2, ceil_mode: bool = False) -> None: ...
def forward(self, input: torch.Tensor) -> torch.Tensor: ...
class EdgeAwareBlurPool2D(nn.Module):
def __init__(self, kernel_size: tuple[int, int] | int,
edge_threshold: float = 1.25,
edge_dilation_kernel_size: int = 3) -> None: ...
def forward(self, input: torch.Tensor, epsilon: float = 1e-6) -> torch.Tensor: ...
def blur_pool2d(input: torch.Tensor, kernel_size: tuple[int, int] | int,
stride: int = 2) -> torch.Tensor: ...
def max_blur_pool2d(input: torch.Tensor, kernel_size: tuple[int, int] | int,
stride: int = 2, max_pool_size: int = 2,
ceil_mode: bool = False) -> torch.Tensor: ...
def edge_aware_blur_pool2d(input: torch.Tensor, kernel_size: tuple[int, int] | int,
edge_threshold: float = 1.25,
edge_dilation_kernel_size: int = 3,
epsilon: float = 1e-6) -> torch.Tensor: ...
Import
from kornia.filters import blur_pool2d, max_blur_pool2d, edge_aware_blur_pool2d
from kornia.filters import BlurPool2D, MaxBlurPool2D, EdgeAwareBlurPool2D
I/O Contract
Inputs (blur_pool2d)
| Name |
Type |
Required |
Description
|
| input |
torch.Tensor (B, C, H, W) |
Yes |
The input feature map to blur and downsample.
|
| kernel_size |
tuple[int, int] or int |
Yes |
Size of the Pascal (binomial) blur kernel.
|
| stride |
int |
No (default 2) |
Stride for the downsampling convolution.
|
Inputs (max_blur_pool2d)
| Name |
Type |
Required |
Description
|
| input |
torch.Tensor (B, C, H, W) |
Yes |
The input feature map.
|
| kernel_size |
tuple[int, int] or int |
Yes |
Size of the blur kernel.
|
| stride |
int |
No (default 2) |
Stride for pooling.
|
| max_pool_size |
int |
No (default 2) |
Kernel size for max pooling applied before blur.
|
| ceil_mode |
bool |
No (default False) |
Whether to use ceil mode in max pooling.
|
Inputs (edge_aware_blur_pool2d)
| Name |
Type |
Required |
Description
|
| input |
torch.Tensor (B, C, H, W) |
Yes |
The input image to blur while preserving edges.
|
| kernel_size |
tuple[int, int] or int |
Yes |
Size of the Gaussian blur kernel.
|
| edge_threshold |
float |
No (default 1.25) |
Positive threshold for edge detection.
|
| edge_dilation_kernel_size |
int |
No (default 3) |
Kernel size for dilating the edge map.
|
| epsilon |
float |
No (default 1e-6) |
Small value for numerical stability.
|
Outputs
| Name |
Type |
Description
|
| output |
torch.Tensor |
The blurred and downsampled feature map. Shape depends on stride and kernel_size: (B, C, H_out, W_out) for blur_pool2d/max_blur_pool2d, or (B, C, H, W) for edge_aware_blur_pool2d (stride=1 internally).
|
Usage Examples
import torch
from kornia.filters import blur_pool2d, max_blur_pool2d, BlurPool2D
# Functional blur pool
input = torch.eye(5)[None, None] # (1, 1, 5, 5)
output = blur_pool2d(input, kernel_size=3, stride=2)
print(output.shape) # torch.Size([1, 1, 3, 3])
# Module-based blur pool
bp = BlurPool2D(kernel_size=3, stride=2)
output = bp(input)
# Max blur pool: equivalent to nn.MaxPool2d + BlurPool2D
output = max_blur_pool2d(input, kernel_size=3, stride=2, max_pool_size=2)
Related Pages
