Exploiting AirSim as a Cross-dataset Benchmark for Safe UAV Landing

and Monocular Depth Estimation Models

Jon Ander I

niguez De Gordoa

, Javier Barandiaran

and Marcos Nieto

Vicomtech Foundation, Basque Research and Technology Alliance (BRTA)

Mikeletegi 57, 20009 Donostia-San Sebasti

an, Spain

Keywords:

Monocular Depth Estimation, Safe Drone Landing, UAV, Synthetic Dataset, Simulation.

Abstract:

As there is a lack of publicly available datasets with depth and surface normal information from a drone’s

view, in this paper, we introduce the synthetic and photorealistic AirSimNC dataset. This dataset is used as

a benchmark to test the zero-shot cross-dataset performance of monocular depth and safe drone landing area

estimation models. We analysed state-of-the-art Deep Learning networks and trained them on the SafeUAV

dataset. While the depth models achieved very satisfactory results in the SafeUAV dataset, they showed a

scaling error in the AirSimNC benchmark. We also compared the performance of networks trained on the

KITTI and NYUv2 datasets, in order to test how training the networks on a bird’s eye view affects in the

performance on our benchmark. Regarding the safe landing estimation models, they surprisingly showed

barely any zero-shot cross-dataset penalty when it comes to the precision of horizontal surfaces.

1 INTRODUCTION

Autonomous drones must retrieve information from

their surroundings in order to ensure a safe ﬂight and

landing. On the one hand, embedded safety mecha-

nisms based on ultrasound sensors are limited to short

range applications. On the other hand, long distance

safety mechanisms, such as LiDAR, are expensive

and too heavy for small UAVs to carry them on board.

Consequently, novel UAV models are trying to exploit

Deep Learning and Computer Vision techniques as an

alternative in order to get environment information.

While depth maps and surface normals can be es-

timated from RGB images using Deep Learning tech-

niques, most depth and surface normal datasets that

are publicly available have been directed towards au-

tonomous driving or indoor applications. However,

there is a lack of open-source and high resolution 3D

datasets from bird’s eye view. Generating a dataset

with such characteristics is challenging, not only be-

cause of the technical difﬁculty of retrieving data

from the viewpoint of a drone, but also due to the

safety laws and airspace regulations in each country.

Some approaches such as the SafeUAV dataset tried to

https://orcid.org/0000-0002-9008-5620

https://orcid.org/0000-0002-8135-0410

https://orcid.org/0000-0001-9879-0992

generate a Google Earth-based semi-synthetic depth

and landing area dataset (Marcu et al., 2019), but

the reconstructed images are quite coarse and some-

times distorted. Therefore, the is no public bench-

mark on which UAV-based Deep Learning models can

be properly compared.

The main contribution of this paper is the exploita-

tion of the AirSim simulator (Shah et al., 2017) by

generating the AirSimNC dataset, which is used as

a benchmark in order to compare the zero-shot cross

dataset performance of different monocular depth and

safe landing area estimation models from a drone

view. This dataset contains photorealistic images

from bird’s eye view at different heights, as well as

the ground truth depth map and information about

safe landing areas. The dataset is also diverse in ef-

fects such as motion blur or different weather con-

ditions. Regarding monocular depth estimation, we

analysed the state-of-the-art AdaBins (Bhat et al.,

2020) and DenseDepth (Alhashim and Wonka, 2018)

networks. We trained them on the SafeUAV dataset

and tested them on the AirSimNC benchmark. We

also tested the same networks trained on the big-

ger KITTI and NYUv2 datasets, in order to check

whether if training the networks on a bird’s eye view

dataset such as SafeUAV (even if is not as photo-

realistic as KITTI or NYUv2) leads to better re-

sults in the AirSimNC benchmark. Regarding the

454

Iñiguez De Gordoa, J., Barandiaran, J. and Nieto, M.

Exploiting AirSim as a Cross-dataset Benchmark for Safe UAV Landing and Monocular Depth Estimation Models.

DOI: 10.5220/0011562200003332

In Proceedings of the 14th International Joint Conference on Computational Intelligence (IJCCI 2022), pages 454-462

ISBN: 978-989-758-611-8; ISSN: 2184-3236

safe landing area estimation models, we analysed the

SafeUAV-Net-Large and Small semantic segmenta-

tion networks, also trained in SafeUAV and tested

them on the AirSimNC dataset.

The rest of this paper is organized as follows. Sec-

tion 2 presents the related work. Section 3 introduces

the six estimation models that were fully or partially

trained on the SafeUAV dataset in order to obtain ro-

bust models that are specialized in bird’s eye view

(Marcu et al., 2019). Section 4 presents the novel Air-

SimNC dataset, and the experimental results obtained

by the different estimation models. The depth esti-

mation models introduced in Section 3 are compared

to state-of-the-art monocular models in Section 4 too.

Section 5 summarizes the most relevant conclusions

of this paper.

2 RELATED WORK

2.1 Deep Learning-based Normal

Estimation

Normal estimation consists of inferring the local ori-

entation of the surfaces from RGB images. Deep

Learning-based approaches have achieved state-of-

the-art performance in surface normal predictions but

they are limited due to the shortage of large and real-

istic outdoor datasets with ground truth information.

Surface normal estimation is relevant to UAV applica-

tions as it can be used for safe landing area estimation.

While some approaches tried to estimate the sur-

face normals using depth map predictions (Yang et al.,

2018), it has been proved that estimating the surface

normals independently leads to a better normal pre-

diction performance (Zhan et al., 2019). Eigen et

al. (Eigen and Fergus, 2015) estimated the surface

normals using a multi-scale convolutional architec-

ture in order to estimate the normal vector for each

image pixel. Li et al. (Li et al., 2015) combined a

deep CNN with a post-proccessing reﬁning step based

on Conditional Random Fields. Wang et al. (Wang

et al., 2015) developed architectures that integrated

local and global information from the input image, as

well as information used by classical normal estima-

tion methods. The GeoNet++ model, introduced by

Qi et al. (Qi et al., 2020), incorporates a two-stream

backbone neural network with an edge-aware reﬁne-

ment module for precise boundary detailed outputs.

2.2 Deep Learning-based Monocular

Depth Estimation

Estimating the depth map from just one view is an

ill-posed problem because many 3D scenes can have

the same picture representation on the image plane.

However, monocular methods require a much cheaper

hardware, less computational complexity and no ex-

ternal calibrations or rectiﬁcations, at the expense of

less accurate results compared to stereo methods.

Many current monocular approaches share a sim-

ilar encoder-decoder architecture. The most popu-

lar encoders are EfﬁcientNets (Bhat et al., 2020),

ResNets (Laina et al., 2016), or variants of ResNet

such as ResNext (Kim et al., 2020; Yin et al., 2019)

or DenseNet (Alhashim and Wonka, 2018; Lee et al.,

2019). For the decoder, Laina et al. (Laina et al.,

2016) proposes an up-scaling decoder architecture

based on up-convolution blocks, while Lee et al. (Lee

et al., 2019) uses Local Planar Guidance layers. Skip

connections are also popular in monocular depth ar-

chitectures, which are usually followed by an atten-

tion mechanism (Kim et al., 2020), or a dilated resid-

ual block (Yin et al., 2019). Aich et al. introduced

an architecture with bidirectional attention modules

(Aich et al., 2021). Moreover, Zhang et al. ex-

ploits temporal consistency among consecutive video

frames (Zhang et al., 2019).

2.3 Synthetic Datasets

Retrieving fully annotated ground truth from real

scenes can be very expensive and complicated, spe-

cially from bird’s eye view. Some existing datasets

such as Okutama-Action (Barekatain et al., 2017) or

VisDrone2018 (Zhu et al., 2018) are based on real

drone view footage but only contain object detection

ground truth (i.e. no depth maps or landing areas).

Moreover, the publicly available depth and normal

datasets either have sparse ground truth or employ

colorization methods. Synthetic datasets are an attrac-

tive alternative for dense and accurate ground truth

generation. There are different approaches for syn-

thetic dataset generation: from open-source 3D ani-

mated ﬁlms (Butler et al., 2012; Mayer et al., 2016),

to vehicle simulators such as CARLA (Sekkat et al.,

2022) or AirSim. The AirSim simulator has already

been used in order to generate the VirtualCity3D (Liu

and Huang, 2021) and UrbanScene3D datasets (Lin

et al., 2022), which contain ground truth information

about bounding boxes, instance segmentation or 3D

pointclouds.

Exploiting AirSim as a Cross-dataset Benchmark for Safe UAV Landing and Monocular Depth Estimation Models

455

3 TRAINING on SafeUAV

DATASET

Due to the absence of comparable and publicly

available sets, we used the semi-synthetic SafeUAV

dataset to train all the networks analysed in this

paper. This dataset comprises RGB images from

3D-reconstructed urban and suburban areas retrieved

from Google Earth, their corresponding dense depth

maps and labelled images with the orientation of each

surface. Pixels are labelled as ’horizontal’, ’vertical’

or ’other’, and this task is referred to as HVO seg-

mentation. This labelling is an oversimpliﬁcation of

the world, since not every horizontal surface is actu-

ally safe for drone landing (e.g. water or roads).

In order to generate monocular depth and HVO es-

timation models for UAV applications, different net-

works were trained on the SafeUAV dataset.

3.1 Studied Deep Learning-based

Architectures

The monocular depth estimation architectures ana-

lyzed in this work are AdaBins and DenseDepth. Ad-

aBins divides a given depth range into N adaptive

bins, and at the time of writing this paper, it presented

the best compromise among the evaluation metrics

in the KITTI Depth and NYUv2 benchmarks (Bhat

et al., 2020). On the other hand, DenseDepth shows

a strong compromise with its relatively low computa-

tional complexity and a marginal performance penalty

(Alhashim and Wonka, 2018).

The landing area estimation networks analyzed

in this paper were proposed in (Marcu et al., 2019).

SafeUAV-Net-Large (or Large model), is presented as

the model with the higher accuracy, while SafeUAV-

Net-Small (or the Small model) is presented as ther

faster and computationally less expensive model.

3.2 Implementation Details

We trained all of our networks on a single NVIDIA

Tesla V100-SXM2 GPU with 32 GB memory. To

avoid over-ﬁtting, we used data augmentation tech-

niques. We used the Adam Optimizer (Kingma and

Ba, 2015), with 40 epochs, a learning rate of 0.0001,

patience of 3 and a batch size of 8.

AdaBins got the best results in the KITTI and

NYUv2 benchmarks when the number of bins was set

to 256. We trained the AdaBins network for 256 and

80 bins, generating the AdaBins-256 and AdaBins-80

models, respectively.

Since DenseDepth showed worse results than Ad-

aBins on the same benchmarks, we analyzed if its per-

formance could be enhanced by training it on mixed

datasets. We trained the DenseDepth network on the

SafeUAV dataset, and then in a super-dataset contain-

ing the SafeUAV and NYUv2 datasets. This way, we

generated two models: DD-UAV and DD-UAV-NYU.

Table 1 shows the size and inferring speed of the

models. The speed values should only be consid-

ered as a reference to compare the complexity of each

model. There are different tools that allow to reduce

the inferring time of networks that have not been used

on these models yet.

3.3 Experimental Results of Depth

Models

Table 2 shows the evaluation metrics of our depth es-

timation models on the SafeUAV dataset. Overall,

the AdaBins-256 models obtained better results than

AdaBins-80. DD-UAV-NYU got considerably better

error metrics than DD-UAV. In order to obtain a wider

visualization of the performance of our models, we

calculated two more parameters, as shown in Table

3: the structural similarity SSIM (Wang et al., 2004)

and the median relative value (Median Rel) between

the estimated and ground truth depth maps. The Me-

dian Rel indicates scaling errors in the estimations,

and it is more robust against outliers than other mean

error metrics. Table 3 shows that both DD-UAV and

DD-UAV-NYU have a slight scaling error (the esti-

mated depth values are 7.2% and 3.6% shorter than

the ground truth depth values, respectively).

3.4 Experimental Results of HVO

Models

Table 4 shows the classiﬁcation evaluation metrics for

the Large and Small models. The row ’Horizontal’

shows the evaluation metrics when the classiﬁcation

problem is reduced to a ’horizontal/non-horizontal’

binary classiﬁcation model. The ’Other’ and ’Verti-

cal’ rows were calculated in a similar way. For safe

landing applications, the precision at predicting hori-

zontal surfaces stands out in importance. Rather than

being able to ﬁnd all the horizontal areas from one

frame, it would be more useful to have a model where

all the predicted horizontal areas were actually hori-

zontal. In that sense, the higher the horizontal pre-

cision is, the safer it will be to use the model for

drone landing applications. In the SafeUAV dataset,

the Large model shows a higher horizontal precision

than the Small model.

ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems

456

Table 1: Number of parameters, size and inference time of each estimation model.

Model Number of parameters (M) Model size (MB) Tesla V100 (images/s)

AdaBins-80 78 M 917 9.623

AdaBins-256 78 M 918 9.583

DD-UAV 42.6 M 337 28.835

DD-UAV-NYU 42.6 M 337 28.835

SafeUAV-Net-Large 24 M 650 27.853

SafeUAV-Net-Small 1 M 24 101.366

Table 2: Quantitative performance of our depth models on SafeUAV. The RMS error is shown in meters. Best results are in

bold, while the second best results are underlined.

Model

Accuracy (higher is better) Error (lower is better)

Abs Rel RMS log

AdaBins-80 0.952 0.987 0.992 0.084 6.425 0.033

AdaBins-256 0.972 0.993 0.996 0.065 5.587 0.026

DD-UAV 0.943 0.998 1 0.130 11.118 0.052

DD-UAV-NYU 0.973 0.996 0.999 0.068 8.451 0.030

Table 3: SSIM and Median Rel of our depth estimation

models on SafeUAV. Best results are in bold, second best

results are underlined.

Model SSIM Median Rel

AdaBins-80 0.331 0.004

AdaBins-256 0.318 -0.009

DD-UAV 0.583 0.072

DD-UAV-NYU 0.524 0.036

4 TESTS on AirSim SIMULATOR

Datasets tend to have strong built-in biases. There-

fore, rather than training and testing models on sub-

sets from the same datasets, a more faithful real-world

performance is obtained by evaluating the networks

on a completely different dataset from the one it was

trained on. This is known as the zero-shot cross-

dataset performance. In order to analyze the cross-

dataset performance of our trained models, we gen-

erated a synthetic dataset using the AirSim simula-

tor. AirSim is an open-source simulator for drones,

cars and other vehicles, and it creates accurate and

real-world environments by taking advantage of the

rendering, physics and perception computation of Un-

real Engine. By exploiting the photorealism of Unreal

Engine and the multiple scene settings that are repro-

ducible with AirSim, we generated a 3D photorealis-

tic dataset which is diverse in weather conditions and

camera effects such as motion blur or noise.

4.1 Synthetic AirSimNC Dataset

Generation

Our synthetic dataset, called AirSimNC, was created

using the suburban AirSimNH and urban CityEnviron

AirSim environments. The synthetic images were re-

trieved from random drone positions with ﬁxed pitch

and roll angles of 45°and 0°, respectively, and a uni-

formly randomized elevation between 30 and 90 me-

ters. From each perspective, RGB, depth and nor-

mal information was retrieved. In order to generate

a semantic segmentation map that classiﬁes the im-

age surfaces as horizontal, vertical or other, the nor-

mal maps were retrieved using the simulator. Then,

the normal vector was computed for each pixel in the

normal map. Afterwards, the angle between the hori-

zontal plane and the normal vector was calculated for

each pixel. Finally, each pixel was classiﬁed as hori-

zontal, vertical or sloped surface by applying empiri-

cally set thresholds into its corresponding angle.

4.2 Experimental Results of Depth

Models

We analyzed the performance of the four depth

estimation models introduced on Section 3. We

also evaluated the performance of the AdaBins and

DenseDepth networks when they are trained on the

KITTI and NYUv2 datasets.

The ﬁrst half of Table 5 shows the evaluation

metrics of the models in the AirSimNC dataset (The

log

error could not be calculated for the DD-KITTI

and DD-NYU models because of undeﬁned terms).

The estimations suffer from a strong scaling error,

as shown in the Median Rel values of Table 6 (in

Exploiting AirSim as a Cross-dataset Benchmark for Safe UAV Landing and Monocular Depth Estimation Models

457

Table 4: Classiﬁcation metrics of the HVO models on SafeUAV. Row ’Average’ shows the average of the metrics in the three

classes. Row ’Reported’ shows the metrics claimed by (Marcu et al., 2019).

SafeUAV Accuracy Precision Recall IoU

Large model

Horizontal 0.827 0.756 0.725 0.587

Other 0.725 0.674 0.751 0.551

Vertical 0.852 0.678 0.559 0.442

Average 0.801 0.703 0.678 0.527

Reported 0.846 0.761 0.748 0.607

Small model

Horizontal 0.817 0.703 0.804 0.600

Other 0.733 0.715 0.675 0.532

Vertical 0.860 0.689 0.610 0.478

Average 0.804 0.702 0.696 0.537

Reported 0.823 0.728 0.693 0.551

Table 5: Quantitative performance of different depth models on the AirSimNC dataset. Best results in each category are in

bold, while the second best results are underlined. Rows with an asterisk represent the metrics for each model, once the

estimated depth maps were scaled to the ground truth median depth.

Model

Accuracy (higher is better) Error (lower is better)

Abs Rel RMS log

AdaBins-80 0.388 0.623 0.804 0.608 42.592 0.191

AdaBins-256 0.249 0.504 0.724 0.930 51.104 0.233

AdaBins-256-KITTI 0.291 0.539 0.725 0.615 52.469 0.218

AdaBins-256-NYU 0.166 0.339 0.519 1.409 70.165 0.309

DD-UAV 0.240 0.396 0.507 0.530 107.630 0.383

DD-UAV-NYU 0.223 0.444 0.647 0.447 60.149 0.249

DD-KITTI 0.208 0.407 0.585 1.041 63.518 -

DD-NYU 0.091 0.190 0.302 3.114 76.365 -

* AdaBins-80 0.588 0.802 0.900 0.338 33.442 0.124

* AdaBins-256 0.567 0.801 0.899 0.393 33.893 0.128

* AdaBins-256-KITTI 0.435 0.745 0.848 0.387 42.254 0.156

* AdaBins-256-NYU 0.415 0.688 0.840 0.380 46.670 0.164

* DD-UAV 0.496 0.762 0.891 0.338 97.095 0.135

* DD-UAV-NYU 0.532 0.808 0.920 0.293 52.056 0.123

* DD-KITTI 0.239 0.459 0.649 0.958 61.548 -

* DD-NYU 0.258 0.492 0.681 0.837 21.257 -

our AdaBins-80-UAV model, for example, the esti-

mated depths are 27.3% higher than the ground truth

depths, while the depth maps estimated by our DD-

UAV models are 66.2% shorter than the real depth

maps).

In a real-world application, a model could be cal-

ibrated in a well-known environment in order to di-

minish the scaling error. To simulate such calibra-

tion, we multiplied each estimated depth map with a

scaling factor, so that the median of the scaled depth

map matches the median of the ground truth (which is

similar to the procedure explained in (Alhashim and

Wonka, 2018)). The second half of Table 5 shows

the evaluation metrics of the depth estimation mod-

els, once the scaling correction is performed.

Our four depth models achieved better scaled re-

ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems

458

Table 7: Evaluation metrics of the Large and Small models on the synthetic AirSimNC dataset.

AirSimNC Accuracy Precision Recall IoU

Large model

Horizontal 0.664 0.724 0.483 0.404

Other 0.567 0.252 0.625 0.215

Vertical 0.700 0.467 0.281 0.219

Average 0.640 0.481 0.463 0.279

Small model

Horizontal 0.634 0.738 0.366 0.324

Other 0.558 0.251 0.629 0.220

Vertical 0.673 0.379 0.273 0.197

Average 0.622 0.456 0.423 0.247

Table 6: SSIM and Median Rel metrics of depth models on

the AirSimNC dataset. Best results are in bold, second best

results are underlined.

Model SSIM Median Rel

AdaBins-80 0.099 -0.175

AdaBins-256 0.092 -0.564

AdaBins-256-KITTI 0.178 -0.125

AdaBins-256-NYU 0.057 -1.185

DD-UAV 0.121 0.524

DD-UAV-NYU 0.147 0.205

DD-KITTI 0.002 -0.044

DD-NYU -0.001 -1.880

sults than the other four models in almost every met-

ric, even if the later models were trained on much

larger and realistic datasets. The best evaluation met-

rics were obtained by the (scaled) DD-UAV-NYU

model in most parameters, followed by AdaBins-80.

This shows how monocular depth models can beneﬁt

from mixing depth datasets during the training pro-

cess.

Figure 1 shows estimations of our depth models,

compared to the ground truth depth. The estimations

in AirSimNH (second column) suffer from outliers:

in high-frequency objects in the AdaBins models, and

in the upper part of the depth maps in the DenseDepth

estimations.

4.3 Experimental Results of HVO

Models

Table 7 shows the evaluation metrics for the HVO

models on the AirSimNC dataset. For both models,

the precision for horizontal surfaces is higher than

72%, with barely no cross-dataset penalty. (the Small

model shows a slightly higher precision). Out of the

estimated ’horizontal’ pixels that were incorrectly la-

belled, most belonged to the ’other’ class.

Figure 2 shows a visual comparison of the HVO

estimation models. The number of ’other’ pixels is

higher in the estimations than in the ground truth (the

HVO models have learned to ’play it safe’ around un-

certain or borderline surfaces). The Large model pro-

vides more detailed estimations (the edges of surfaces

and objects are more discernible in the Large estima-

tions), while the estimations of the Small models are

more vague overall.

5 CONCLUSION

This paper introduced the synthetic and photoreal-

istic AirSimNC dataset, which was employed as a

benchmark to compare the cross-dataset performance

of depth and HVO estimation models. Regarding

the depth estimation models, the ones trained on

SafeUAV beat other state-of-the-art models in our

benchmark. However, we also checked that mix-

ing a semi-synthetic dataset with bird’s eye view

(SafeUAV) and a real dataset with more realistic im-

ages (NYUv2) leads to a great improvement in the

performance of the DenseDepth network. Overall,

the models trained of bird’s eye view obtained decent

results in our benchmark, up to a scaling error. Re-

garding the safe landing area estimation models, they

showed barely any cross-dataset penalty at the preci-

sion of horizontal surfaces.

During this paper, we have regarded depth and

landing area estimation as independent tasks. The ge-

ometric relationship between depth and surface nor-

Exploiting AirSim as a Cross-dataset Benchmark for Safe UAV Landing and Monocular Depth Estimation Models

459

Figure 1: Qualitative comparison of our depth estimation models. First row: RGB images from SafeUAV, AirSimNH and

CityEnviron, respectively. Second row: ground truth depth maps. Rows 3 to 6: estimations by the AdaBins-80, AdaBins-256,

DD-UAV and DD-UAV-NYU.

ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems

460

Figure 2: Qualitative comparison of the HVO estimation models. First row: RGB images from SafeUAV, AirSimNH and

CityEnviron, respectively. Second row: ground truth HVO labels. Rows 3 to 4: Large and Small model predictions.

mals could be exploited so that a single model per-

forms both estimation tasks simultaneously. The de-

velopment of such network is left as future work.

ACKNOWLEDGEMENTS

This work has received funding from Basque Gov-

ernment under project CODISAVA of the program

ELKARTEK (ETORTEK)-2020.

REFERENCES

Aich, S., Uwabeza Vianney, J. M., Amirul Islam, M., and

Bingbing Liu, M. K. (2021). Bidirectional attention

network for monocular depth estimation. In 2021

IEEE International Conference on Robotics and Au-

tomation (ICRA), pages 11746–11752.

Alhashim, I. and Wonka, P. (2018). High quality monocular

depth estimation via transfer learning. arXiv e-prints,

abs/1812.11941.

Barekatain, M., Marti, M., Shih, H., Murray, S., Nakayama,

K., Matsuo, Y., and Prendinger, H. (2017). Okutama-

action: An aerial view video dataset for concurrent

human action detection. In 2017 IEEE Conference

on Computer Vision and Pattern Recognition Work-

shops (CVPRW), pages 2153–2160, Los Alamitos,

CA, USA. IEEE Computer Society.

Bhat, S. F., Alhashim, I., and Wonka, P. (2020). Adabins:

Depth estimation using adaptive bins. arXiv, pages

1–13.

Butler, D. J., Wulff, J., Stanley, G. B., and Black, M. J.

(2012). A naturalistic open source movie for optical

ﬂow evaluation. Lecture Notes in Computer Science

(including subseries Lecture Notes in Artiﬁcial Intel-

Exploiting AirSim as a Cross-dataset Benchmark for Safe UAV Landing and Monocular Depth Estimation Models

461

ligence and Lecture Notes in Bioinformatics), 7577

LNCS(PART 6):611–625.

Eigen, D. and Fergus, R. (2015). Predicting depth, surface

normals and semantic labels with a common multi-

scale convolutional architecture. Proceedings of the

IEEE International Conference on Computer Vision,

2015 Inter:2650–2658.

Kim, D., Lee, S., Lee, J., and Kim, J. (2020). Leveraging

Contextual Information for Monocular Depth Estima-

tion. IEEE Access, 8:147808–147817.

Kingma, D. P. and Ba, J. L. (2015). Adam: A method for

stochastic optimization. 3rd International Conference

on Learning Representations, ICLR 2015 - Confer-

ence Track Proceedings, pages 1–15.

Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., and

Navab, N. (2016). Deeper depth prediction with fully

convolutional residual networks. Proceedings - 2016

4th International Conference on 3D Vision, 3DV 2016,

pages 239–248.

Lee, J. H., Han, M.-K., Ko, D. W., and Suh, I. H. (2019).

From big to small: Multi-scale local planar guid-

ance for monocular depth estimation. arXiv preprint

arXiv:1907.10326.

Li, B., Shen, C., Dai, Y., van den Hengel, A., and He, M.

(2015). Depth and surface normal estimation from

monocular images using regression on deep features

and hierarchical crfs. In 2015 IEEE Conference on

Computer Vision and Pattern Recognition (CVPR),

pages 1119–1127.

Lin, L., Liu, Y., Hu, Y., Yan, X., Xie, K., and Huang, H.

(2022). Capturing, reconstructing, and simulating: the

urbanscene3d dataset. In ECCV.

Liu, Y. and Huang, H. (2021). Virtualcity3d: A large scale

urban scene dataset and simulator.

Marcu, A., Costea, D., Lic

aret¸, V., P

ırvu, M., Slus¸anschi,

E., and Leordeanu, M. (2019). SafeUAV: learning to

estimate depth and safe landing areas for UAVs from

synthetic data. Lecture Notes in Computer Science

(including subseries Lecture Notes in Artiﬁcial Intel-

ligence and Lecture Notes in Bioinformatics), 11130

LNCS:43–58.

Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D.,

Dosovitskiy, A., and Brox, T. (2016). A Large Dataset

to Train Convolutional Networks for Disparity, Opti-

cal Flow, and Scene Flow Estimation. Proceedings of

the IEEE Computer Society Conference on Computer

Vision and Pattern Recognition, 2016-Decem:4040–

4048.

Qi, X., Liu, Z., Liao, R., Torr, P. H., Urtasun, R., and Jia, J.

(2020). GeoNet++: Iterative Geometric Neural Net-

work with Edge-Aware Reﬁnement for Joint Depth

and Surface Normal Estimation. IEEE Transactions

on Pattern Analysis and Machine Intelligence, pages

1–1.

Sekkat, A. R., Dupuis, Y., Kumar, V. R., Rashed, H., Yo-

gamani, S., Vasseur, P., and Honeine, P. (2022). Syn-

woodscape: Synthetic surround-view ﬁsheye camera

dataset for autonomous driving. IEEE Robotics and

Automation Letters, 7(3):8502–8509.

Shah, S., Dey, D., Lovett, C., and Kapoor, A. (2017). Air-

sim: High-ﬁdelity visual and physical simulation for

autonomous vehicles. In Field and Service Robotics.

Wang, X., Fouhey, D. F., and Gupta, A. (2015). Designing

deep networks for surface normal estimation. Pro-

ceedings of the IEEE Computer Society Conference

on Computer Vision and Pattern Recognition, 07-12-

June(Figure 2):539–547.

Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P.

(2004). Image quality assessment: From error visi-

bility to structural similarity. IEEE Transactions on

Image Processing, 13(4):600–612.

Yang, Z., Wang, P., Xu, W., Zhao, L., and Neva-

tia, R. (2018). Unsupervised learning of geometry

from videos with edge-aware depth-normal consis-

tency. 32nd AAAI Conference on Artiﬁcial Intelli-

gence, AAAI 2018, 1(c):7493–7500.

Yin, W., Liu, Y., Shen, C., and Yan, Y. (2019). Enforcing

geometric constraints of virtual normal for depth pre-

diction. Proceedings of the IEEE International Con-

ference on Computer Vision, 2019-Octob:5683–5692.

Zhan, H., Weerasekera, C. S., Garg, R., and Reid, I. (2019).

Self-supervised learning for single view depth and

surface normal estimation. Proceedings - IEEE In-

ternational Conference on Robotics and Automation,

2019-May:4811–4817.

Zhang, H., Shen, C., Li, Y., Cao, Y., Liu, Y., and Yan,

Y. (2019). Exploiting temporal consistency for real-

time video depth estimation. Proceedings of the IEEE

International Conference on Computer Vision, 2019-

Octob:1725–1734.

Zhu, P., Wen, L., Bian, X., Ling, H., and Hu, Q. (2018).

Vision Meets Drones: A Challenge. pages 1–11.

ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems

462