FisheyeSuperPoint: Keypoint Detection and Description Network for
Fisheye Images
Anna Konrad
, Ciar
an Eising
, Ganesh Sistu
, John McDonald
, Rudi Villing
and Senthil Yogamani
Hamilton Institute, Maynooth University, Ireland
Department of Electronic Engineering, Maynooth University, Ireland
Department of Electronic & Computer Engineering, University of Limerick, Ireland
Valeo Vision Systems, Galway, Ireland
Department of Computer Science, Maynooth University, Ireland
Keypoints, Interest Points, Feature Detection, Feature Description, Fisheye Images, Deep Learning.
Keypoint detection and description is a commonly used building block in computer vision systems particularly
for robotics and autonomous driving. However, the majority of techniques to date have focused on standard
cameras with little consideration given to fisheye cameras which are commonly used in urban driving and
automated parking. In this paper, we propose a novel training and evaluation pipeline for fisheye images.
We make use of SuperPoint as our baseline which is a self-supervised keypoint detector and descriptor that
has achieved state-of-the-art results on homography estimation. We introduce a fisheye adaptation pipeline
to enable training on undistorted fisheye images. We evaluate the performance on the HPatches benchmark,
and, by introducing a fisheye based evaluation method for detection repeatability and descriptor matching
correctness, on the Oxford RobotCar dataset.
Keypoint detection and description is a fundamental
step in computer vision for image registration (Ma
et al., 2021). It has a wide range of applications in-
cluding 3D reconstruction, object tracking, video sta-
bilization and SLAM. Approaches designed to be in-
variant to changes in scale, illumination, perspective,
etc. have been extensively studied in the computer vi-
sion literature. However, despite their prevalence in
automotive and robotic systems, few approaches have
explicitly considered fisheye images that pose an ad-
ditional challenge of spatially variant distortion. Thus
a patch in the centre of an image looks different com-
pared to a region in the periphery of the image where
the radial distortion is much higher. Fisheye cameras
are a fundamental sensor in autonomous driving nec-
essary to cover the near field around the vehicle (Yo-
gamani et al., 2019). Four fisheye cameras on each
side of the vehicle can cover the entire 360
field of
A common approach to use fisheye images for
tasks in computer vision is to first rectify the image
and then apply algorithms suited for standard images
(Lo et al., 2018), (Esparza et al., 2014). A principal
drawback of such methods is that the field of view
is reduced and resampling artifacts can be introduced
into the periphery of the image. An alternative ap-
proach, as demonstrated in (Ravi Kumar et al., 2020),
is to work directly in the fisheye image space and
thereby avoid such issues. Recently, there has been
progress with such approaches for various visual per-
ception tasks such as dense matching (H
ane et al.,
2014), object detection (Rashed et al., 2021), depth
estimation (Kumar et al., 2018), re-localisation (Tri-
pathi and Yogamani, 2021), soiling detection (Uric
et al., 2019) and people detection (Duan et al., 2020).
Feature detectors and descriptors can describe cor-
ners (also called interest points or keypoints), edges or
morphological region features. Traditionally, feature
detection and description has been done with hand-
crafted algorithms (Ma et al., 2021). Some of the
most well-known algorithms include Harris (Harris
and Stephens, 1998), FAST (Rosten and Drummond,
2006) and SIFT (Lowe, 2004). Thorough reviews
of traditional and modern techniques have been con-
ducted in various surveys (Ma et al., 2021), (Mikola-
jczyk and Schmid, 2005), (Li et al., 2015).
Konrad, A., Eising, C., Sistu, G., McDonald, J., Villing, R. and Yogamani, S.
FisheyeSuperPoint: Keypoint Detection and Description Network for Fisheye Images.
DOI: 10.5220/0010795400003124
In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 4: VISAPP, pages
ISBN: 978-989-758-555-5; ISSN: 2184-4321
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
Recently, several CNN based feature correspon-
dence techniques have been explored which outper-
form classical features. For example, a universal cor-
respondence network in (Sarlin et al., 2016) demon-
strates state-of-the-art results on various datasets by
making use of a spatial transformer to normalise for
affine transformations. This is an example of feature
correspondence learning independent of the applica-
tion in which it is used. It is an open problem to learn
feature correspondence which is optimal for the later
stages in the perception pipeline e.g. bundle adjust-
ment. For instance, end-to-end learning of feature
matching could possibly learn diversity and distribu-
tion rather than focusing solely on measures such as
distinctiveness and repeatability. This is particularly
useful for training on fisheye images which have a do-
main gap compared to regular images. In addition,
learning based detectors permit the encoder backend
to be merged into a multi-task model. Such multi-task
models share the same encoder, which can improve
performance and reduce latency for all tasks (Leang
et al., 2020), (Ravikumar et al., 2021).
SuperPoint is a self-supervised CNN framework
for keypoint detection and description (DeTone et al.,
2018). It consists of one encoder and two different
decoders for the detection and the description output.
It is pretrained as an corner detector via a syntheti-
cally generated dataset containing basic shapes like
rectangles, lines, stars, etc. The resulting corner de-
tector shows inconsistent keypoint detections on the
same scene for varying camera viewpoints. To im-
prove on the detection consistency for varying cam-
era viewpoints, homographic adaptation is used to
generate a superset of keypoints from random homo-
graphic warpings on the MS-COCO training dataset
(Lin et al., 2014). The network is then trained on
those keypoints and the whole process is repeated for
several iterations. Recently, it has achieved state-of-
the-art results in several benchmarks in combination
with the SuperGlue matcher (Sarlin et al., 2020).
We adapt the SuperPoint feature detector and de-
scriptor so it can be trained and evaluated on fisheye
images. The contributions of this paper include:
Random fisheye warping and unwarping in place
of homographic transforms.
Implementation of fisheye adaption for self-
supervised training of the SuperPoint network.
Evaluation of the repeatability of detectors and
comparison of their performance to FisheyeSu-
perPoint on fisheye and standard images.
Evaluation of the matching correctness of de-
scriptors and comparison of their performance to
FisheyeSuperPoint on fisheye & standard images.
Figure 1: Example imagery of the Oxford RobotCar Dataset
(Maddern et al., 2017), showing different weather and light-
ing conditions.
We propose FisheyeSuperPoint, a SuperPoint net-
work that has been trained self-supervised on fisheye
images. To enable this, the homographic adaptation
step in the training pipeline has been exchanged with
our fisheye adaptation process to cope with the non-
linear mapping of fisheye images.
2.1 Datasets Used
The data used for training of the FisheyeSuperPoint
network is a subset of the Oxford RobotCar Dataset
(Maddern et al., 2017) (RobotCar). The RobotCar
dataset contains fisheye images, stereo images, LI-
DAR sensor readings and GPS from a vehicle across
various seasons and weather conditions. Some exam-
ples from the fisheye images are shown in Figure 1.
The data is made available in subsets for each drive.
The subsets contain data from three different fisheye
cameras which were mounted on the left, right and
rear of the vehicle. To generate a representative sub-
set of fisheye images, image sequences from each of
seven different weather conditions (sun, clouds, over-
cast, rain, snow, night, dusk) were used with 840k im-
ages in total. To reduce the total number of images
and duplicates in the resulting dataset, we sampled
every 10th frame from the sequences.
The resulting training dataset contained 84k
1024x1024 fisheye images. During training, the im-
ages were downsampled to 256x256 images to match
the resizing of MS-COCO (Lin et al., 2014) that was
used for the training of SuperPoint (DeTone et al.,
2.2 Fisheye Warping and Unwarping
In order to train FisheyeSuperPoint on fisheye im-
ages, a substitute for the homographic warping in Su-
perPoint is needed. The pinhole camera model is
the standard projection function for much research
in computer vision, as in SuperPoint (DeTone et al.,
2018). The pinhole projection function is given as
FisheyeSuperPoint: Keypoint Detection and Description Network for Fisheye Images
Figure 2: Results of the Fisheye Warping. a) the original image, b) the warped image, c) the unwarped image and d) the
difference between the original and unwarped image. c) is the result of applying the inverse of the fisheye warping to b). The
structural differences in d) are regions which were outside of the warped image b). The remaining differences in d) are due to
the bilinear interpolation used for warping and unwarping the image.
p =
f X
, where X = (X ,Y, Z)
is a point in
the camera coordinate system, and f is the nominal
focal length of the pinhole camera.
Fisheye functions provide a nonlinear mapping
from the camera coordinate system (e.g. Figure 2).
We can define a mapping from R
to the fisheye im-
age as
π : R
A true inverse is naturally not possible, as all
depth information is lost in the formation of the im-
age. However, we can define an unprojection map-
ping from the fisheye image domain to the unit central
projective sphere:
: I
Unfortunately, the Oxford RobotCar Dataset does
not provide details of the fisheye model they use, nor
its parameters. However, they provide a look-up-table
that can map from a distorted to an undistorted im-
age. We use this look-up-table and fit the fourth order
polynomial model p(θ) described below.
In principle, it does not matter exactly which fish-
eye mapping function is used, as long as it provides
a reasonably accurate model of the image transforma-
tion. In our case, we use a radial polynomial function
for π, as per (Yogamani et al., 2019):
π(X) =
, d =
p(θ) = a
θ + a
+ . . . + a
θ = arccos
+ Z
where p(θ) is a polynomial of order n, with n = 4
typically sufficient.
In SuperPoint (DeTone et al., 2018), random ho-
mographies are used to simulate multiple camera
viewpoints. In order to train a SuperPoint network
with fisheye images, the homographic warping needs
to be replaced with an equivalent transform that is ap-
plicable to fisheye imagery. Using the fisheye func-
tions (π and π
) described above, we can consider
the following steps to generate a new fisheye warped
1. Each point is projected from the image I
to a unit
sphere S
using π
2. A new virtual camera position is selected by a ran-
dom rotation R and translation t (six degrees of
freedom Euclidean transform)
3. Each point on the unit sphere is reprojected to this
new, virtual camera position, by applying π(RX +
This results in a mapping from I
, where
represents the new image with the random Eu-
clidean transform applied. We call this mapping fish-
eye warping F and fisheye unwarping F
F (I
) = π(Rπ
) + t) (2)
In practice, to avoid sparsity in the new image,
each pixel on the warped image is inverse transformed
to the corresponding sub-pixel location on the original
image and sampled using bilinear interpolation. Ad-
ditionally, as the fisheye unprojection function π
computationally costly due to the requirement for a
polynomial root solver, a set of 2000 look-up-tables
is pre-computed for the image warpings. For each
look-up-table, we sample the rotation and translation
along each axis uniformly at random from the inter-
val U
, 30
] and U
[0.3, 0.3] relative to
the unit sphere S
, respectively. In order to invert
this warping, as required for unwarping the detected
point responses, the inverse of the steps 1 - 3 above
are followed. An example of the resulting images is
shown in Figure 2. The original image a) is warped to
b) and unwarped to c) using the pre-computed look-
up-tables. The difference d) between the original a)
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications
Figure 3: Self-supervised fisheye Keypoint Detection and Description training framework adapted from SuperPoint frame-
work (DeTone et al., 2018). Random fisheye warpings are applied to a single image. Point responses are received from the
base detector when applying it to the warped images. The point responses are then unwarped and accumulated to get an
aggregated keypoint superset of the original image.
and unwarped image c) shows no structural differ-
ences apart from the regions which were outside of
the warped image b) and therefore remain black in the
unwarped image. The other differences are due to the
bilinear interpolation used for warping and unwarping
the image.
2.3 Fisheye Adaptation
The fisheye warping is incorporated into the Super-
Point training pipeline as shown in Figure 3. The aim
of the adaptation process is to provide consistent key-
point responses on the same scene under varying cam-
era viewpoints. A single fisheye image I
and a base
keypoint detector f
are input to the fisheye adapta-
tion process. Within that process, a random fisheye
warping F is sampled and applied to the image with
bilinear interpolation resulting in a warped fisheye
image I
. The base detector is used to detect point
responses x
, which are subsequently unwarped back
to the original image space by applying the inverse
fisheye warping F
The adaptation process is repeated for N
= 100
random fisheye warpings and the resulting point re-
sponses are accumulated. The point responses from
different fisheye warpings are expected to differ ini-
tially. By accumulating the point responses into a su-
perset of detected keypoints, a new ground-truth for
keypoints in the original image is generated. Subse-
quently, the base detector is trained on this superset
to generate a superior, more consistent keypoint de-
tector. The full process can be repeated iteratively in
order to enhance performance and consistency of the
resulting model. The descriptor decoder is learned in
a semi-dense manner in the last iteration of the model
enhancement, as described in (DeTone et al., 2018).
FisheyeSuperPoint is developed on top of a train-
able tensorflow implementation (Pautrat and Sarlin,
2021) based on SuperPoint (DeTone et al., 2018). To
train FisheyeSuperPoint, we use a magic-point net-
work (Pautrat and Sarlin, 2021) trained on MS-COCO
as a base detector, where we apply fisheye adaptation
to the RobotCar dataset and train FisheyeSuperPoint
with 600,000 iterations. To train SuperPoint, we use
the same pretrained network, where we apply homo-
graphic adaptation to the MS-COCO dataset and train
SuperPoint with 600,000 iterations.
The training including two fisheye/homographic
adaptation iterations is executed on two Nvidia GTX
1080Ti GPUs and takes approximately one week to
complete. The duration of the training with fish-
eye adaptation is similar to the training with homo-
graphic adaptation in SuperPoint. While the fisheye
adaptation is a more complex procedure, the use of
pre-calculated look-up-tables for the fisheye warping
means it is computationally efficient. Note that the
same testing data is used for all models in the experi-
3.1 Benchmark Setup
HPatches: The performance of FisheyeSuperPoint
and SuperPoint in comparison to traditional corner
detection techniques is evaluated by the repeatabil-
ity of detections and the homography estimation cor-
rectness on the HPatches benchmark (Balntas et al.,
2017). It contains multiple images of planar objects
from varying camera viewpoints or with different il-
luminations. As the ground truth homography is pro-
vided for each corresponding image pair, detections
FisheyeSuperPoint: Keypoint Detection and Description Network for Fisheye Images
on an image pair can be warped and their consis-
tency can be compared. The detection repeatability
of FisheyeSuperPoint and SuperPoint is compared to
the detection algorithms FAST (Rosten and Drum-
mond, 2006), Harris (Harris and Stephens, 1998) and
Shi (Shi and Tomasi, 1994). The homography estima-
tion correctness of FisheyeSuperPoint and SuperPoint
is compared to SIFT (Lowe, 2004) and ORB (Rublee
et al., 2011). The evaluation methodology is the same
as described in (DeTone et al., 2018), where the ho-
mography is estimated with OpenCV based on nearest
neighbour matching on the descriptors.
The estimated homography is compared to the
ground truth homography of HPatches by transform-
ing four corner points c
of the image with both ho-
mographies, resulting in c
and ˆc
with j = 4. As
per (DeTone et al., 2018), the homography estimation
correctness is then calculated based on the distance
between the corner points by
|| ε
and averaged over all n = 295 test images.
Fisheye Oxford RobotCar: In order to assess the
performance of the trained networks on fisheye im-
ages, we evaluate keypoint detection repeatability and
matching correctness using a fisheye test set. Unfortu-
nately, there currently is no fisheye image equivalent
to the HPatches benchmark available that contains
precise ground truth viewpoint relations. Therefore,
to evaluate performance on fisheye images, we gener-
ate an artificial dataset based on a test set of RobotCar
(Maddern et al., 2017). The test set was generated
from approximately 51k images across five different
weather conditions and image sequences that had not
been used in the training set for FisheyeSuperPoint.
To increase diversity in the test set, one out of every
172 frames was sampled resulting in a base test set of
300 images.
From the base test set, 300 illumination change
test images are created by applying gamma correc-
tion, with gamma for each image drawn randomly
from a uniform distribution γ [0.1, 2]. Indepen-
dently, each image of the base test set is warped to
create a viewpoint change test image using a fisheye
warping F from a set of 300 random fisheye warpings
not previously used for training.
By applying the same keypoint detector to corre-
sponding images we generate two different point re-
x = f
), x
= f
In the case of viewpoint changes, point responses that
lie outside the overlapping region of both images are
filtered by applying boolean masks. The masks are
generated by warping all-ones matrices of image size
with F and F
for x
and x, respectively. We ap-
ply F (x) to warp the point response x into the image
space of x
and calculate the detector repeatability as
described in (DeTone et al., 2018).
In order to evaluate the descriptor matching cor-
rectness M
on RobotCar, nearest neighbour match-
ing is performed on the descriptors, resulting in a
set of matches M. The keypoints x
of the resulting
matches in I
are warped using F (x
) and the eu-
clidean distance d() to their corresponding keypoint,
, is calculated. The inliers G are calculated as
G = {x
, x
: d(x
, x
) < ε}. The threshold with
ε = 3px is set to the same distance as for the homog-
raphy estimation correctness. The matching correct-
ness M
is calculated as the ratio of inliers to matches
, where |G| denotes the number of elements
in the set G. We average the descriptor matching cor-
rectness M
over all n = 300 test images for one given
In addition, we report the root-mean-square error
(RMSE) of the distance in all i = n ×k matches by
+ d
+ ···+ d
for one given model. The number of keypoint detec-
tions k = 300 and the number of test images n = 300
is kept constant for all experiments.
3.2 Comparison
A comparison of the detection repeatability on
FisheyeSuperPoint, SuperPoint, as well as FAST
(Rosten and Drummond, 2006), Harris (Harris and
Stephens, 1998) and Shi (Shi and Tomasi, 1994) is
shown in Table 1. The default OpenCV implementa-
tion is used for FAST, Harris and Shi. We apply Non-
Maximum Suppression (NMS) on a square mask with
size of 4px, 8px to the keypoint detections. The num-
ber of detected points k = 300 and correct distance
threshold ε = 3 stays constant. Images in HPatches
are resized to 240 ×320 and RobotCar images are re-
sized to 256 ×256.
FisheyeSuperPoint outperforms the other detec-
tors for viewpoint and illumination changes in Robot-
Car when a high NMS is applied. While the tradi-
tional Harris detector outperforms FisheyeSuperPoint
with a NMS = 8 on viewpoint changes, it ranks sec-
ond and is superior to SuperPoint, FAST and Shi.
For the detection repeatability on the HPatches
benchmark, FisheyeSuperPoint outperforms the clas-
sical detectors on the scenes with illumination
changes. For the scenes with viewpoint changes, it
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications
Table 1: Detector Repeatability on HPatches and RobotCar.
Algorithm Illumination Changes Viewpoint Changes
FisheyeSuperPoint 0.664 0.631 0.678 0.626
SuperPoint (DeTone et al., 2018) 0.663 0.622 0.672 0.610
FAST (Rosten and Drummond, 2006) 0.576 0.493 0.598 0.492
Harris (Harris and Stephens, 1998) 0.630 0.590 0.725 0.612
Shi (Shi and Tomasi, 1994) 0.584 0.515 0.613 0.523
FisheyeSuperPoint 0.896 0.876 0.768 0.716
SuperPoint (DeTone et al., 2018) 0.897 0.869 0.754 0.708
FAST (Rosten and Drummond, 2006) 0.837 0.751 0.724 0.569
Harris (Harris and Stephens, 1998) 0.876 0.841 0.827 0.693
Shi (Shi and Tomasi, 1994) 0.831 0.769 0.709 0.604
Table 2: Homography correctness on HPatches, descriptor matching correctness and RMSE on RobotCar.
Algorithm HPatches H
RobotCar M
RobotCar RMSE
FisheyeSuperPoint 0.712 0.862 38.4
SuperPoint (DeTone et al., 2018) 0.668 0.859 36.3
SIFT (Lowe, 2004) 0.766 0.663 120.7
ORB (Rublee et al., 2011) 0.414 0.463 136.9
outperforms the other detectors when a higher NMS
is applied. FisheyeSuperPoint and SuperPoint consis-
tently outperform FAST and Shi. While the repeata-
bility values for all detectors are lower in (DeTone
et al., 2018) the ranking of the detectors is consistent.
The results match the values reported by (Pautrat and
Sarlin, 2021).
The results of the homography and matching cor-
rectness on HPatches and RobotCar are shown in Ta-
ble 2. In addition to FisheyeSuperPoint and Super-
Point, we compare the performance to SIFT (Lowe,
2004) and ORB (Rublee et al., 2011) which are imple-
mented using OpenCV. NMS = 8 and ε = 3 is applied
for all experiments. The number of detected points is
set to k = 1000 for HPatches and k = 300 for Robot-
Car. The HPatches images are resized to 480 ×640
and the RobotCar images to 512 ×512.
Both FisheyeSuperPoint and SuperPoint show a
superior performance compared to SIFT and ORB
when used for descriptor matching on the RobotCar
test data. This is particularly evident when taking
into account the RMSE results as per 3, which in-
dicate a high number of outliers for SIFT and Orb
with RMSE > 100. The matching performance on
RobotCar is also shown in Figure 4, where matches
with a euclidean distance of d 3px are indicated
with green lines. SIFT outperforms the other algo-
rithms for the homography correctness on HPatches.
FisheyeSuperPoint ranks second and is superior to
SuperPoint and ORB.
This work describes the new FisheyeSuperPoint key-
point detection and description network which uses
a pipeline to train and evaluate it directly on fisheye
image datasets. To enable the self-supervised training
on fisheye images, fisheye warping is utilised. The
fisheye image is mapped to a new, warped fisheye im-
age through the intermediate step of projection to a
unit sphere, with the camera’s virtual pose being var-
ied in six degrees of freedom. This process is embed-
ded in an existing SuperPoint implementation (Pautrat
and Sarlin, 2021) and trained on the RobotCar dataset
(Maddern et al., 2017).
In order to compare the performance of Fisheye-
SuperPoint to other detectors, we introduce a method
to evaluate keypoint detection repeatability and
matching correctness on fisheye images. Fisheye-
SuperPoint consistently outperforms SuperPoint for
the experiments on standard images (HPatches), es-
pecially in terms of homography correctness. This
might be due to more variations in the RobotCar train-
ing data. Both FisheyeSuperPoint and SuperPoint
perform similarly in our fisheye evaluations. This
was unexpected and hints towards the robustness of
the SuperPoint network, suggesting that it could be
used for keypoint detection and description on fisheye
images directly. Further evaluations on non-artificial
data for descriptor matching correctness could pro-
vide a better insight into the performance of both net-
FisheyeSuperPoint: Keypoint Detection and Description Network for Fisheye Images
Figure 4: Qualitative results of feature matching on RobotCar images with k = 300 detected points. Nearest neighbour
matches with a distance d 3px are shown in green. From left to right: FisheyeSuperPoint, SuperPoint (DeTone et al., 2018),
SIFT (Lowe, 2004), ORB (Rublee et al., 2011).
While Harris and SIFT achieve higher repeatabil-
ities and homography correctness than FisheyeSuper-
Point in a few cases, our method comes with several
advantages. One of those advantages is the adapt-
ability of the network structure, which could be en-
hance with alternatives such as deformable convolu-
tional layers (Dai et al., 2017). The adaptive man-
ner of deformable convolutional layers could help in-
crease detection and description performance under
the influence of the radial distortion in fisheye images.
Another opportunity is to incorporate FisheyeSu-
perPoint into multi-task visual perception networks
like Omnidet (Ravikumar et al., 2021). Multi-task
networks can present advantages in computational
complexity and performance by sharing base layers
of a network, which will be enhanced with Fisheye-
This publication has emanated from research con-
ducted with the financial support of Science Founda-
tion Ireland (SFI) under Grant number 18/CRT/6049
and 16/RI/3399. The opinions, findings and conclu-
sions or recommendations expressed in this material
are those of the author(s) and do not necessarily re-
flect the views of the Science Foundation Ireland.
Balntas, V., Lenc, K., Vedaldi, A., and Mikolajczyk, K.
(2017). HPatches: A benchmark and evaluation of
handcrafted and learned local descriptors. In Proceed-
ings of CVPR, pages 5173–5182.
Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., and
Wei, Y. (2017). Deformable convolutional networks.
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications
In Proceedings of the IEEE International Conference
on Computer Vision (ICCV).
DeTone, D., Malisiewicz, T., and Rabinovich, A. (2018).
Superpoint: Self-supervised interest point detection
and description. In Proceedings of the CVPR Work-
shops (CVPR-W), pages 224–236.
Duan, Z., Tezcan, O., Nakamura, H., Ishwar, P., and Kon-
rad, J. (2020). Rapid: rotation-aware people detec-
tion in overhead fisheye images. In Proceedings of
the CVPR Workshops, pages 636–637.
Esparza, J., Helmle, M., and J
ahne, B. (2014). Wide base
stereo with fisheye optics: A robust approach for 3d
reconstruction in driving assistance. In Jiang, X.,
Hornegger, J., and Koch, R., editors, Pattern Recog-
nition, pages 342–353, Cham. Springer International
Harris, C. and Stephens, M. (1998). A combined corner
and edge detector. In Alvey vision conference, pages
ane, C., Heng, L., Lee, G. H., Sizov, A., and Pollefeys,
M. (2014). Real-time direct dense matching on fish-
eye images using plane-sweeping stereo. In 2014 2nd
International Conference on 3D Vision, volume 1.
Kumar, V. R., Milz, S., Witt, C., Simon, M., Amende,
K., Petzold, J., Yogamani, S., and Pech, T. (2018).
Near-field depth estimation using monocular fisheye
camera: A semi-supervised learning approach using
sparse lidar data. In Proceedings of the CVPR Work-
shops (CVPR-W), volume 7.
Leang, I., Sistu, G., B
urger, F., Bursuc, A., and Yoga-
mani, S. (2020). Dynamic task weighting methods
for multi-task networks in autonomous driving sys-
tems. In 2020 IEEE 23rd International Conference on
Intelligent Transportation Systems (ITSC), pages 1–8.
Li, Y., Wang, S., Tian, Q., and Ding, X. (2015). A survey
of recent advances in visual feature detection. Neuro-
computing, 149.
Lin, T., Maire, M., Belongie, S. J., Bourdev, L. D., Girshick,
R. B., Hays, J., Perona, P., Ramanan, D., Doll
ar, P.,
and Zitnick, C. L. (2014). Microsoft COCO: common
objects in context. In Proceedings of the European
Conference on Computer Vision (ECCV), pages 740–
Lo, I., Shih, K., and Chen, H. H. (2018). Image stitching
for dual fisheye cameras. In 25th IEEE International
Conference on Image Processing (ICIP), pages 3164–
Lowe, D. G. (2004). Distinctive image features from scale-
invariant keypoints. International Journal of Com-
puter Vision, 60:91–110.
Ma, J., Jiang, X., Fan, A., Jiang, J., and Yan, J. (2021).
Image matching from handcrafted to deep features:
A survey. International Journal of Computer Vision,
Maddern, W., Pascoe, G., Linegar, C., and Newman, P.
(2017). 1 year, 1000 km: The oxford robotcar
dataset. The International Journal of Robotics Re-
search, 36(1):3–15.
Mikolajczyk, K. and Schmid, C. (2005). A perfor-
mance evaluation of local descriptors. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence,
Pautrat, R. and Sarlin, P.-E. (2021). Github SuperPoint Im-
Rashed, H., Mohamed, E., Sistu, G., Kumar, V. R., Eis-
ing, C., El-Sallab, A., and Yogamani, S. (2021). Gen-
eralized object detection on fisheye cameras for au-
tonomous driving: Dataset, representations and base-
line. In Proceedings of the IEEE/CVF WACV.
Ravi Kumar, V., Yogamani, S., Bach, M., Witt, C.,
Milz, S., and Mader, P. (2020). UnRectDepthNet:
Self-Supervised Monocular Depth Estimation using a
Generic Framework for Handling Common Camera
Distortion Models. In International Conference on In-
telligent Robots and Systems (IROS).
Ravikumar, V., Yogamani, S., Rashed, H., Sistu, G., Witt,
C., Leang, I., Milz, S., and Mader, P. (2021). Om-
nidet: Surround view cameras based multi-task visual
perception network for autonomous driving. IEEE
Robotics and Automation Letters, pages 1–1.
Rosten, E. and Drummond, T. (2006). Machine learning
for high-speed corner detection. In Proceedings of the
European Conference on Computer Vision (ECCV).
Rublee, E., Rabaud, V., Konolige, K., and Bradski, G.
(2011). Orb: An efficient alternative to sift or surf. In
Proceedings of the IEEE/CVF International Confer-
ence on Computer Vision (ICCV), pages 2564–2571.
Sarlin, P.-E., DeTone, D., Malisiewicz, T., and Rabinovich,
A. (2016). Universal correspondence network. In Pro-
ceedings of Conference on Neural Information Pro-
cessing Systems (NIPS).
Sarlin, P.-E., DeTone, D., Malisiewicz, T., and Rabinovich,
A. (2020). Superglue: Learning feature matching
with graph neural networks. In Proceedings of CVPR,
pages 4938–4947.
Shi, J. and Tomasi, C. (1994). Good features to track. In
Proceedings of CVPR.
Tripathi, N. and Yogamani, S. (2021). Trained trajectory
based automated parking system using Visual SLAM.
In Proceedings of CVPR Workshops.
ar, M., Ulicny, J., Sistu, G., Rashed, H., Krizek, P.,
Hurych, D., Vobecky, A., and Yogamani, S. (2019).
Desoiling dataset: Restoring soiled areas on automo-
tive fisheye cameras. In Proceedings of the IEEE/CVF
International Conference on Computer Vision Work-
Yogamani, S., Hughes, C., Horgan, J., Sistu, G., Varley,
P., O’Dea, D., Uric
ar, M., et al. (2019). Wood-
scape: A multi-task, multi-camera fisheye dataset for
autonomous driving. In Proceedings of the IEEE/CVF
International Conference on Computer Vision (ICCV).
FisheyeSuperPoint: Keypoint Detection and Description Network for Fisheye Images