import math
from typing import List, Optional, Tuple
import torch
import torch.nn.functional as F
from typing_extensions import TypedDict
from kornia.core import Device, Module, Tensor, concatenate, eye, tensor, where, zeros
from kornia.core.check import KORNIA_CHECK_SHAPE
from kornia.geometry.subpix import ConvSoftArgmax3d, NonMaximaSuppression2d
from kornia.geometry.transform import ScalePyramid, pyrdown, resize
from .laf import laf_from_center_scale_ori, laf_is_inside_image
from .orientation import PassLAF
from .responses import BlobHessian
def _scale_index_to_scale(max_coords: Tensor, sigmas: Tensor, num_levels: int) -> Tensor:
r"""Auxiliary function for ScaleSpaceDetector. Converts scale level index from ConvSoftArgmax3d to the actual
scale, using the sigmas from the ScalePyramid output.
Args:
max_coords: tensor [BxNx3].
sigmas: tensor [BxNxD], D >= 1
Returns:
tensor [BxNx3].
"""
# depth (scale) in coord_max is represented as (float) index, not the scale yet.
# we will interpolate the scale using pytorch.grid_sample function
# Because grid_sample is for 4d input only, we will create fake 2nd dimension
# ToDo: replace with 3d input, when grid_sample will start to support it
# Reshape for grid shape
B, N, _ = max_coords.shape
scale_coords = max_coords[:, :, 0].contiguous().view(-1, 1, 1, 1)
# Replace the scale_x_y
out = concatenate(
[sigmas[0, 0] * torch.pow(2.0, scale_coords / float(num_levels)).view(B, N, 1), max_coords[:, :, 1:]], 2
)
return out
def _create_octave_mask(mask: Tensor, octave_shape: List[int]) -> Tensor:
r"""Downsample a mask based on the given octave shape."""
mask_shape = octave_shape[-2:]
mask_octave = F.interpolate(mask, mask_shape, mode="bilinear", align_corners=False)
return mask_octave.unsqueeze(1)
[docs]class ScaleSpaceDetector(Module):
r"""Module for differentiable local feature detection, as close as possible to classical local feature detectors
like Harris, Hessian-Affine or SIFT (DoG).
It has 5 modules inside: scale pyramid generator, response ("cornerness") function,
soft nms function, affine shape estimator and patch orientation estimator.
Each of those modules could be replaced with learned custom one, as long, as
they respect output shape.
Args:
num_features: Number of features to detect. In order to keep everything batchable,
output would always have num_features output, even for completely homogeneous images.
mr_size: multiplier for local feature scale compared to the detection scale.
6.0 is matching OpenCV 12.0 convention for SIFT.
scale_pyr_module: generates scale pyramid. See :class:`~kornia.geometry.ScalePyramid` for details.
Default: ScalePyramid(3, 1.6, 10).
resp_module: calculates ``'cornerness'`` of the pixel.
nms_module: outputs per-patch coordinates of the response maxima.
See :class:`~kornia.geometry.ConvSoftArgmax3d` for details.
ori_module: for local feature orientation estimation. Default:class:`~kornia.feature.PassLAF`,
which does nothing. See :class:`~kornia.feature.LAFOrienter` for details.
aff_module: for local feature affine shape estimation. Default: :class:`~kornia.feature.PassLAF`,
which does nothing. See :class:`~kornia.feature.LAFAffineShapeEstimator` for details.
minima_are_also_good: if True, then both response function minima and maxima are detected
Useful for symmetric response functions like DoG or Hessian. Default is False
"""
def __init__(
self,
num_features: int = 500,
mr_size: float = 6.0,
scale_pyr_module: Module = ScalePyramid(3, 1.6, 15),
resp_module: Module = BlobHessian(),
nms_module: Module = ConvSoftArgmax3d(
(3, 3, 3), (1, 1, 1), (1, 1, 1), normalized_coordinates=False, output_value=True
),
ori_module: Module = PassLAF(),
aff_module: Module = PassLAF(),
minima_are_also_good: bool = False,
scale_space_response: bool = False,
) -> None:
super().__init__()
self.mr_size = mr_size
self.num_features = num_features
self.scale_pyr = scale_pyr_module
self.resp = resp_module
self.nms = nms_module
self.ori = ori_module
self.aff = aff_module
self.minima_are_also_good = minima_are_also_good
# scale_space_response should be True if the response function works on scale space
# like Difference-of-Gaussians
self.scale_space_response = scale_space_response
def __repr__(self) -> str:
return (
f"{self.__class__.__name__}("
f"num_features={self.num_features}, "
f"mr_size={self.mr_size}, "
f"scale_pyr={self.scale_pyr.__repr__()}, "
f"resp={self.resp.__repr__()}, "
f"nms={self.nms.__repr__()}, "
f"ori={self.ori.__repr__()}, "
f"aff={self.aff.__repr__()})"
)
def detect(self, img: Tensor, num_feats: int, mask: Optional[Tensor] = None) -> Tuple[Tensor, Tensor]:
dev: Device = img.device
dtype: torch.dtype = img.dtype
sigmas: List[Tensor]
sp, sigmas, _ = self.scale_pyr(img)
all_responses: List[Tensor] = []
all_lafs: List[Tensor] = []
px_size = 0.5 if self.scale_pyr.double_image else 1.0
for oct_idx, octave in enumerate(sp):
sigmas_oct = sigmas[oct_idx]
B, CH, L, H, W = octave.size()
# Run response function
if self.scale_space_response:
oct_resp = self.resp(octave, sigmas_oct.view(-1))
else:
oct_resp = self.resp(octave.permute(0, 2, 1, 3, 4).reshape(B * L, CH, H, W), sigmas_oct.view(-1)).view(
B, L, CH, H, W
)
# We want nms for scale responses, so reorder to (B, CH, L, H, W)
oct_resp = oct_resp.permute(0, 2, 1, 3, 4)
# 3rd extra level is required for DoG only
if isinstance(self.scale_pyr.extra_levels, Tensor) and self.scale_pyr.extra_levels % 2 != 0:
oct_resp = oct_resp[:, :, :-1]
if mask is not None:
oct_mask: Tensor = _create_octave_mask(mask, oct_resp.shape)
oct_resp = oct_mask * oct_resp
# Differentiable nms
coord_max: Tensor
response_max: Tensor
coord_max, response_max = self.nms(oct_resp)
if self.minima_are_also_good:
coord_min, response_min = self.nms(-oct_resp)
take_min_mask = (response_min > response_max).to(response_max.dtype)
response_max = response_min * take_min_mask + (1 - take_min_mask) * response_max
coord_max = coord_min * take_min_mask.unsqueeze(2) + (1 - take_min_mask.unsqueeze(2)) * coord_max
# Now, lets crop out some small responses
responses_flatten = response_max.view(response_max.size(0), -1) # [B, N]
max_coords_flatten = coord_max.view(response_max.size(0), 3, -1).permute(0, 2, 1) # [B, N, 3]
if responses_flatten.size(1) > num_feats:
resp_flat_best, idxs = torch.topk(responses_flatten, k=num_feats, dim=1)
max_coords_best = torch.gather(max_coords_flatten, 1, idxs.unsqueeze(-1).repeat(1, 1, 3))
else:
resp_flat_best = responses_flatten
max_coords_best = max_coords_flatten
B, N = resp_flat_best.size()
# Converts scale level index from ConvSoftArgmax3d to the actual scale, using the sigmas
if isinstance(self.scale_pyr.n_levels, Tensor):
num_levels = int(self.scale_pyr.n_levels.item())
elif isinstance(self.scale_pyr.n_levels, int):
num_levels = self.scale_pyr.n_levels
else:
raise TypeError(
"Expected the scale pyramid module to have `n_levels` as a Tensor or int."
f"Gotcha {type(self.scale_pyr.n_levels)}"
)
max_coords_best = _scale_index_to_scale(max_coords_best, sigmas_oct, num_levels)
# Create local affine frames (LAFs)
rotmat = eye(2, dtype=dtype, device=dev).view(1, 1, 2, 2)
current_lafs = concatenate(
[
self.mr_size * max_coords_best[:, :, 0].view(B, N, 1, 1) * rotmat,
max_coords_best[:, :, 1:3].view(B, N, 2, 1),
],
3,
)
# Zero response lafs, which touch the boundary
good_mask = laf_is_inside_image(current_lafs, octave[:, 0])
resp_flat_best = resp_flat_best * good_mask.to(dev, dtype)
# Normalize LAFs
current_lafs *= px_size
all_responses.append(resp_flat_best)
all_lafs.append(current_lafs)
px_size *= 2
# Sort and keep best n
responses = concatenate(all_responses, 1)
lafs = concatenate(all_lafs, 1)
responses, idxs = torch.topk(responses, k=num_feats, dim=1)
lafs = torch.gather(lafs, 1, idxs.unsqueeze(-1).unsqueeze(-1).repeat(1, 1, 2, 3))
return responses, lafs
[docs] def forward(self, img: Tensor, mask: Optional[Tensor] = None) -> Tuple[Tensor, Tensor]:
"""Three stage local feature detection. First the location and scale of interest points are determined by
detect function. Then affine shape and orientation.
Args:
img: image to extract features with shape [BxCxHxW]
mask: a mask with weights where to apply the response function. The shape must be the same as
the input image.
Returns:
lafs: shape [BxNx2x3]. Detected local affine frames.
responses: shape [BxNx1]. Response function values for corresponding lafs
"""
responses, lafs = self.detect(img, self.num_features, mask)
lafs = self.aff(lafs, img)
lafs = self.ori(lafs, img)
return lafs, responses
class Detector_config(TypedDict):
nms_size: int
pyramid_levels: int
up_levels: int
scale_factor_levels: float
s_mult: float
def get_default_detector_config() -> Detector_config:
return {
# Extraction Parameters
"nms_size": 15,
"pyramid_levels": 4,
"up_levels": 1,
"scale_factor_levels": math.sqrt(2),
"s_mult": 22.0,
}
[docs]class MultiResolutionDetector(Module):
"""Multi-scale feature detector, based on code from KeyNet. Can be used with any response function.
This is based on the original code from paper
"Key.Net: Keypoint Detection by Handcrafted and Learned CNN Filters".
See :cite:`KeyNet2019` for more details.
Args:
model: response function, such as KeyNet or BlobHessian
num_features: Number of features to detect.
conf: Dict with initialization parameters. Do not pass it, unless you know what you are doing`.
ori_module: for local feature orientation estimation. Default: :class:`~kornia.feature.PassLAF`,
which does nothing. See :class:`~kornia.feature.LAFOrienter` for details.
aff_module: for local feature affine shape estimation. Default: :class:`~kornia.feature.PassLAF`,
which does nothing. See :class:`~kornia.feature.LAFAffineShapeEstimator` for details.
"""
def __init__(
self,
model: Module,
num_features: int = 2048,
config: Detector_config = get_default_detector_config(),
ori_module: Optional[Module] = None,
aff_module: Optional[Module] = None,
) -> None:
super().__init__()
self.model = model
# Load extraction configuration
self.num_pyramid_levels = config["pyramid_levels"]
self.num_upscale_levels = config["up_levels"]
self.scale_factor_levels = config["scale_factor_levels"]
self.mr_size = config["s_mult"]
self.nms_size = config["nms_size"]
self.nms = NonMaximaSuppression2d((self.nms_size, self.nms_size))
self.num_features = num_features
if ori_module is None:
self.ori: Module = PassLAF()
else:
self.ori = ori_module
if aff_module is None:
self.aff: Module = PassLAF()
else:
self.aff = aff_module
[docs] def remove_borders(self, score_map: Tensor, borders: int = 15) -> Tensor:
"""It removes the borders of the image to avoid detections on the corners."""
mask = torch.zeros_like(score_map)
mask[:, :, borders:-borders, borders:-borders] = 1
return mask * score_map
def detect_features_on_single_level(
self, level_img: Tensor, num_kp: int, factor: Tuple[float, float]
) -> Tuple[Tensor, Tensor]:
det_map = self.nms(self.remove_borders(self.model(level_img)))
device = level_img.device
dtype = level_img.dtype
yx = det_map.nonzero()[:, 2:].t()
scores = det_map[0, 0, yx[0], yx[1]] # keynet supports only non-batched images
scores_sorted, indices = torch.sort(scores, descending=True)
indices = indices[where(scores_sorted > 0.0)]
yx = yx[:, indices[:num_kp]].t()
current_kp_num = len(yx)
xy_projected = yx.view(1, current_kp_num, 2).flip(2) * tensor(factor, device=device, dtype=dtype)
scale_factor = 0.5 * (factor[0] + factor[1])
scale = scale_factor * self.mr_size * torch.ones(1, current_kp_num, 1, 1, device=device, dtype=dtype)
lafs = laf_from_center_scale_ori(xy_projected, scale, zeros(1, current_kp_num, 1, device=device, dtype=dtype))
return scores_sorted[:num_kp], lafs
def detect(self, img: Tensor, mask: Optional[Tensor] = None) -> Tuple[Tensor, Tensor]:
# Compute points per level
num_features_per_level: List[float] = []
tmp = 0.0
factor_points = self.scale_factor_levels**2
levels = self.num_pyramid_levels + self.num_upscale_levels + 1
for idx_level in range(levels):
tmp += factor_points ** (-1 * (idx_level - self.num_upscale_levels))
nf = self.num_features * factor_points ** (-1 * (idx_level - self.num_upscale_levels))
num_features_per_level.append(nf)
num_features_per_level = [int(x / tmp) for x in num_features_per_level]
_, _, h, w = img.shape
img_up = img
cur_img = img
all_responses: List[Tensor] = []
all_lafs: List[Tensor] = []
# Extract features from the upper levels
for idx_level in range(self.num_upscale_levels):
nf = num_features_per_level[len(num_features_per_level) - self.num_pyramid_levels - 1 - (idx_level + 1)]
num_points_level = int(nf)
# Resize input image
up_factor = self.scale_factor_levels ** (1 + idx_level)
nh, nw = int(h * up_factor), int(w * up_factor)
up_factor_kpts = (float(w) / float(nw), float(h) / float(nh))
img_up = resize(img_up, (nh, nw), interpolation="bilinear", align_corners=False)
cur_scores, cur_lafs = self.detect_features_on_single_level(img_up, num_points_level, up_factor_kpts)
all_responses.append(cur_scores.view(1, -1))
all_lafs.append(cur_lafs)
# Extract features from the downsampling pyramid
for idx_level in range(self.num_pyramid_levels + 1):
if idx_level > 0:
cur_img = pyrdown(cur_img, factor=self.scale_factor_levels)
_, _, nh, nw = cur_img.shape
factor = (float(w) / float(nw), float(h) / float(nh))
else:
factor = (1.0, 1.0)
num_points_level = int(num_features_per_level[idx_level])
if idx_level > 0 or (self.num_upscale_levels > 0):
nf2 = [num_features_per_level[a] for a in range(0, idx_level + 1 + self.num_upscale_levels)]
res_points = tensor(nf2).sum().item()
num_points_level = int(res_points)
cur_scores, cur_lafs = self.detect_features_on_single_level(cur_img, num_points_level, factor)
all_responses.append(cur_scores.view(1, -1))
all_lafs.append(cur_lafs)
responses = concatenate(all_responses, 1)
lafs = concatenate(all_lafs, 1)
if lafs.shape[1] > self.num_features:
responses, idxs = torch.topk(responses, k=self.num_features, dim=1)
lafs = torch.gather(lafs, 1, idxs[..., None, None].expand(-1, -1, 2, 3))
return responses, lafs
[docs] def forward(self, img: Tensor, mask: Optional[Tensor] = None) -> Tuple[Tensor, Tensor]:
"""Three stage local feature detection. First the location and scale of interest points are determined by
detect function. Then affine shape and orientation.
Args:
img: image to extract features with shape [1xCxHxW]. KeyNetDetector does not support batch processing,
because the number of detections is different on each image.
mask: a mask with weights where to apply the response function. The shape must be the same as
the input image.
Returns:
lafs: shape [1xNx2x3]. Detected local affine frames.
responses: shape [1xNx1]. Response function values for corresponding lafs
"""
KORNIA_CHECK_SHAPE(img, ["1", "C", "H", "W"])
responses, lafs = self.detect(img, mask)
lafs = self.aff(lafs, img)
lafs = self.ori(lafs, img)
return lafs, responses