"""
# %BANNER_BEGIN%
# ---------------------------------------------------------------------
# %COPYRIGHT_BEGIN%
#
# Magic Leap, Inc. ("COMPANY") CONFIDENTIAL
#
# Unpublished Copyright (c) 2020
# Magic Leap, Inc., All Rights Reserved.
#
# NOTICE: All information contained herein is, and remains the property
# of COMPANY. The intellectual and technical concepts contained herein
# are proprietary to COMPANY and may be covered by U.S. and Foreign
# Patents, patents in process, and are protected by trade secret or
# copyright law. Dissemination of this information or reproduction of
# this material is strictly forbidden unless prior written permission is
# obtained from COMPANY. Access to the source code contained herein is
# hereby forbidden to anyone except current COMPANY employees, managers
# or contractors who have executed Confidentiality and Non-disclosure
# agreements explicitly covering such access.
#
# The copyright notice above does not evidence any actual or intended
# publication or disclosure of this source code, which includes
# information that is confidential and/or proprietary, and is a trade
# secret, of COMPANY. ANY REPRODUCTION, MODIFICATION, DISTRIBUTION,
# PUBLIC PERFORMANCE, OR PUBLIC DISPLAY OF OR THROUGH USE OF THIS
# SOURCE CODE WITHOUT THE EXPRESS WRITTEN CONSENT OF COMPANY IS
# STRICTLY PROHIBITED, AND IN VIOLATION OF APPLICABLE LAWS AND
# INTERNATIONAL TREATIES. THE RECEIPT OR POSSESSION OF THIS SOURCE
# CODE AND/OR RELATED INFORMATION DOES NOT CONVEY OR IMPLY ANY RIGHTS
# TO REPRODUCE, DISCLOSE OR DISTRIBUTE ITS CONTENTS, OR TO MANUFACTURE,
# USE, OR SELL ANYTHING THAT IT MAY DESCRIBE, IN WHOLE OR IN PART.
#
# %COPYRIGHT_END%
# ----------------------------------------------------------------------
# %AUTHORS_BEGIN%
#
# Originating Authors: Paul-Edouard Sarlin
#
# %AUTHORS_END%
# --------------------------------------------------------------------*/
# %BANNER_END%
Described in:
SuperPoint: Self-Supervised Interest Point Detection and Description,
Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich, CVPRW 2018.
Original code: github.com/MagicLeapResearch/SuperPointPretrainedNetwork
Adapted by Philipp Lindenberger (Phil26AT)
"""
import os.path
import torch
from torch import nn
from gluefactory.models.base_model import BaseModel
from gluefactory.models.utils.misc import pad_and_stack
def simple_nms(scores, radius):
"""Perform non maximum suppression on the heatmap using max-pooling.
This method does not suppress contiguous points that have the same score.
Args:
scores: the score heatmap of size `(B, H, W)`.
radius: an integer scalar, the radius of the NMS window.
"""
def max_pool(x):
return torch.nn.functional.max_pool2d(
x, kernel_size=radius * 2 + 1, stride=1, padding=radius
)
zeros = torch.zeros_like(scores)
max_mask = scores == max_pool(scores)
for _ in range(2):
supp_mask = max_pool(max_mask.float()) > 0
supp_scores = torch.where(supp_mask, zeros, scores)
new_max_mask = supp_scores == max_pool(supp_scores)
max_mask = max_mask | (new_max_mask & (~supp_mask))
return torch.where(max_mask, scores, zeros)
def top_k_keypoints(keypoints, scores, k):
if k >= len(keypoints):
return keypoints, scores
scores, indices = torch.topk(scores, k, dim=0, sorted=True)
return keypoints[indices], scores
def sample_k_keypoints(keypoints, scores, k):
if k >= len(keypoints):
return keypoints, scores
indices = torch.multinomial(scores, k, replacement=False)
return keypoints[indices], scores[indices]
def soft_argmax_refinement(keypoints, scores, radius: int):
width = 2 * radius + 1
sum_ = torch.nn.functional.avg_pool2d(
scores[:, None], width, 1, radius, divisor_override=1
)
ar = torch.arange(-radius, radius + 1).to(scores)
kernel_x = ar[None].expand(width, -1)[None, None]
dx = torch.nn.functional.conv2d(scores[:, None], kernel_x, padding=radius)
dy = torch.nn.functional.conv2d(
scores[:, None], kernel_x.transpose(2, 3), padding=radius
)
dydx = torch.stack([dy[:, 0], dx[:, 0]], -1) / sum_[:, 0, :, :, None]
refined_keypoints = []
for i, kpts in enumerate(keypoints):
delta = dydx[i][tuple(kpts.t())]
refined_keypoints.append(kpts.float() + delta)
return refined_keypoints
# Legacy (broken) sampling of the descriptors
def sample_descriptors(keypoints, descriptors, s):
b, c, h, w = descriptors.shape
keypoints = keypoints - s / 2 + 0.5
keypoints /= torch.tensor(
[(w * s - s / 2 - 0.5), (h * s - s / 2 - 0.5)],
).to(
keypoints
)[None]
keypoints = keypoints * 2 - 1 # normalize to (-1, 1)
args = {"align_corners": True} if torch.__version__ >= "1.3" else {}
descriptors = torch.nn.functional.grid_sample(
descriptors, keypoints.view(b, 1, -1, 2), mode="bilinear", **args
)
descriptors = torch.nn.functional.normalize(
descriptors.reshape(b, c, -1), p=2, dim=1
)
return descriptors
# The original keypoint sampling is incorrect. We patch it here but
# keep the original one above for legacy.
def sample_descriptors_fix_sampling(keypoints, descriptors, s: int = 8):
"""Interpolate descriptors at keypoint locations"""
b, c, h, w = descriptors.shape
keypoints = keypoints / (keypoints.new_tensor([w, h]) * s)
keypoints = keypoints * 2 - 1 # normalize to (-1, 1)
descriptors = torch.nn.functional.grid_sample(
descriptors, keypoints.view(b, 1, -1, 2), mode="bilinear", align_corners=False
)
descriptors = torch.nn.functional.normalize(
descriptors.reshape(b, c, -1), p=2, dim=1
)
return descriptors
class SuperPoint(BaseModel):
default_conf = {
"has_detector": True,
"has_descriptor": True,
"descriptor_dim": 256,
# Inference
"sparse_outputs": True,
"dense_outputs": False,
"nms_radius": 4,
"refinement_radius": 0,
"detection_threshold": 0.005,
"max_num_keypoints": -1,
"max_num_keypoints_val": None,
"force_num_keypoints": False,
"randomize_keypoints_training": False,
"remove_borders": 4,
"legacy_sampling": True, # True to use the old broken sampling
}
required_data_keys = ["image"]
checkpoint_url = "https://github.com/magicleap/SuperGluePretrainedNetwork/raw/master/models/weights/superpoint_v1.pth" # noqa: E501
def _init(self, conf):
self.relu = nn.ReLU(inplace=True)
self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
c1, c2, c3, c4, c5 = 64, 64, 128, 128, 256
self.conv1a = nn.Conv2d(1, c1, kernel_size=3, stride=1, padding=1)
self.conv1b = nn.Conv2d(c1, c1, kernel_size=3, stride=1, padding=1)
self.conv2a = nn.Conv2d(c1, c2, kernel_size=3, stride=1, padding=1)
self.conv2b = nn.Conv2d(c2, c2, kernel_size=3, stride=1, padding=1)
self.conv3a = nn.Conv2d(c2, c3, kernel_size=3, stride=1, padding=1)
self.conv3b = nn.Conv2d(c3, c3, kernel_size=3, stride=1, padding=1)
self.conv4a = nn.Conv2d(c3, c4, kernel_size=3, stride=1, padding=1)
self.conv4b = nn.Conv2d(c4, c4, kernel_size=3, stride=1, padding=1)
if conf.has_detector:
self.convPa = nn.Conv2d(c4, c5, kernel_size=3, stride=1, padding=1)
self.convPb = nn.Conv2d(c5, 65, kernel_size=1, stride=1, padding=0)
for param in self.convPa.parameters():
param.requires_grad = False
for param in self.convPb.parameters():
param.requires_grad = False
if conf.has_descriptor:
self.convDa = nn.Conv2d(c4, c5, kernel_size=3, stride=1, padding=1)
self.convDb = nn.Conv2d(
c5, conf.descriptor_dim, kernel_size=1, stride=1, padding=0
)
self.load_state_dict(torch.load(os.path.join('weights', 'superpoint_v1.pth')))
def _forward(self, data):
image = data["image"]
if image.shape[1] == 3: # RGB
scale = image.new_tensor([0.299, 0.587, 0.114]).view(1, 3, 1, 1)
image = (image * scale).sum(1, keepdim=True)
# Shared Encoder
x = self.relu(self.conv1a(image))
x = self.relu(self.conv1b(x))
x = self.pool(x)
x = self.relu(self.conv2a(x))
x = self.relu(self.conv2b(x))
x = self.pool(x)
x = self.relu(self.conv3a(x))
x = self.relu(self.conv3b(x))
x = self.pool(x)
x = self.relu(self.conv4a(x))
x = self.relu(self.conv4b(x))
pred = {}
if self.conf.has_detector:
# Compute the dense keypoint scores
cPa = self.relu(self.convPa(x))
scores = self.convPb(cPa)
scores = torch.nn.functional.softmax(scores, 1)[:, :-1]
b, c, h, w = scores.shape
scores = scores.permute(0, 2, 3, 1).reshape(b, h, w, 8, 8)
scores = scores.permute(0, 1, 3, 2, 4).reshape(b, h * 8, w * 8)
pred["keypoint_scores"] = dense_scores = scores
if self.conf.has_descriptor:
# Compute the dense descriptors
cDa = self.relu(self.convDa(x))
dense_desc = self.convDb(cDa)
dense_desc = torch.nn.functional.normalize(dense_desc, p=2, dim=1)
pred["descriptors"] = dense_desc
if self.conf.sparse_outputs:
assert self.conf.has_detector and self.conf.has_descriptor
scores = simple_nms(scores, self.conf.nms_radius)
# Discard keypoints near the image borders
if self.conf.remove_borders:
scores[:, : self.conf.remove_borders] = -1
scores[:, :, : self.conf.remove_borders] = -1
if "image_size" in data:
for i in range(scores.shape[0]):
w, h = data["image_size"][i]
scores[i, int(h.item()) - self.conf.remove_borders :] = -1
scores[i, :, int(w.item()) - self.conf.remove_borders :] = -1
else:
scores[:, -self.conf.remove_borders :] = -1
scores[:, :, -self.conf.remove_borders :] = -1
# Extract keypoints
best_kp = torch.where(scores > self.conf.detection_threshold)
scores = scores[best_kp]
# Separate into batches
keypoints = [
torch.stack(best_kp[1:3], dim=-1)[best_kp[0] == i] for i in range(b)
]
scores = [scores[best_kp[0] == i] for i in range(b)]
# Keep the k keypoints with highest score
max_kps = self.conf.max_num_keypoints
# for val we allow different
if not self.training and self.conf.max_num_keypoints_val is not None:
max_kps = self.conf.max_num_keypoints_val
# Keep the k keypoints with highest score
if max_kps > 0:
if self.conf.randomize_keypoints_training and self.training:
# instead of selecting top-k, sample k by score weights
keypoints, scores = list(
zip(
*[
sample_k_keypoints(k, s, max_kps)
for k, s in zip(keypoints, scores)
]
)
)
else:
keypoints, scores = list(
zip(
*[
top_k_keypoints(k, s, max_kps)
for k, s in zip(keypoints, scores)
]
)
)
keypoints, scores = list(keypoints), list(scores)
if self.conf["refinement_radius"] > 0:
keypoints = soft_argmax_refinement(
keypoints, dense_scores, self.conf["refinement_radius"]
)
# Convert (h, w) to (x, y)
keypoints = [torch.flip(k, [1]).float() for k in keypoints]
if self.conf.force_num_keypoints:
keypoints = pad_and_stack(
keypoints,
max_kps,
-2,
mode="random_c",
bounds=(
0,
data.get("image_size", torch.tensor(image.shape[-2:]))
.min()
.item(),
),
)
scores = pad_and_stack(scores, max_kps, -1, mode="zeros")
else:
keypoints = torch.stack(keypoints, 0)
scores = torch.stack(scores, 0)
# Extract descriptors
if (len(keypoints) == 1) or self.conf.force_num_keypoints:
# Batch sampling of the descriptors
if self.conf.legacy_sampling:
desc = sample_descriptors(keypoints, dense_desc, 8)
else:
desc = sample_descriptors_fix_sampling(keypoints, dense_desc, 8)
else:
if self.conf.legacy_sampling:
desc = [
sample_descriptors(k[None], d[None], 8)[0]
for k, d in zip(keypoints, dense_desc)
]
else:
desc = [
sample_descriptors_fix_sampling(k[None], d[None], 8)[0]
for k, d in zip(keypoints, dense_desc)
]
pred = {
"keypoints": keypoints + 0.5,
"descriptors": desc.transpose(-1, -2),
}
if self.conf.dense_outputs:
pred["dense_descriptors"] = dense_desc
return pred
def loss(self, pred, data):
raise NotImplementedError
def metrics(self, pred, data):
raise NotImplementedError