Domain Adaptation in LiDAR Semantic Segmentation by Aligning Class

Distributions

nigo Alonso

, Luis Riazuelo

, Luis Montesano

1,2

and Ana C. Murillo

RoPeRt Group, DIIS - I3A, Universidad de Zaragoza, Spain

Bitbrain, Zaragoza, Spain

Keywords:

Robotics, Autonomous Systems, LiDAR, Deep Learning, Semantic Segmentation, Domain Adaptation.

Abstract:

LiDAR semantic segmentation provides 3D semantic information about the environment, an essential cue for

intelligent systems, such as autonomous vehicles, during their decision making processes. Unfortunately, the

annotation process for this task is very expensive. To overcome this, it is key to ﬁnd models that generalize well

or adapt to additional domains where labeled data is limited. This work addresses the problem of unsupervised

domain adaptation for LiDAR semantic segmentation models. We propose simple but effective strategies to

reduce the domain shift by aligning the data distribution on the input space. Besides, we present a learning-

based module to align the distribution of the semantic classes of the target domain to the source domain. Our

approach achieves new state-of-the-art results on three different public datasets, which showcase adaptation to

three different domains.

1 INTRODUCTION

3D semantic segmentation is an important computer

vision task that provides useful information to each

registered 3D point of the surrounding environment.

It has a wide range of applications and is particularly

important in robotics, since most autonomous sys-

tems require an accurate and robust perception of their

environment. LiDAR (Light Detection And Ranging)

is a frequently used sensor for 3D perception in au-

tonomous vehicles, such as cars or delivery robots,

that provides accurate distance measurements of the

surrounding 3D space. Despite LiDAR broad adop-

tion, recognition systems such as semantic segmen-

tation methods that generalize and perform well for

different LiDAR sensors, vehicle set-ups or environ-

ments remains an unsolved challenging problem. Ex-

isting solutions typically need signiﬁcant amounts of

labeled data to adapt to different domains (Mei et al.,

2019).

In recent years, deep learning methods are achiev-

ing state-of-the-art performance in 3D LiDAR seman-

tic segmentation (Milioto et al., 2019; Alonso et al.,

2020) when trained and evaluated in the same domain.

Nevertheless, generalizing the learned knowledge to

new domains or environments not seen during train-

ing is still a frequent open problem. When applying

existing models on data with a different distribution

Figure 1: Domain Adaptation in LiDAR Semantic Segmen-

tation. Top row: segmentation result of a given model train

on the same source domain (Behley et al., 2019). Bottom

row: segmentation of the same model on a target domain

without adaptation and the improved result applying our

proposed adaptation. Note relevant details missed if we do

not adapt the model (marked with a red/green square).

330

Alonso, I., Riazuelo, L., Montesano, L. and Murillo, A.

Domain Adaptation in LiDAR Semantic Segmentation by Aligning Class Distributions.

DOI: 10.5220/0010610703300337

In Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2021), pages 330-337

ISBN: 978-989-758-522-7

than the training data, i.e., from a different domain,

the performance is considerably degraded (Wang and

Deng, 2018; Tzeng et al., 2017; Vu et al., 2019).

Domain adaptation techniques aim to eliminate or

reduce this drop. Existing works for domain adapta-

tion in semantic segmentation focus on RGB data (Vu

et al., 2019; Zou et al., 2019; Chen et al., 2019; Li

et al., 2019). Most of them try to minimize the dis-

tribution shift between two different domains. Very

few approaches have tackled this problem with Li-

DAR data (Wu et al., 2019), which equally suffers

from the domain shift. RGB data commonly suffers

from variations due to light and weather conditions,

while the most common variations within 3D point

clouds data come from sensor resolution (i.e., sensors

with more laser sweeps generate denser point clouds)

and from the sensor placement (because point clouds

have relative coordinates with respect to the sensor).

Both sensor resolution and placement issues are com-

mon examples that change the data distribution of the

captured 3D point clouds. Coping with these issues

would enable the use of large existing labeled LiDAR

datasets for more realistic use-cases in robotic appli-

cations, reducing the need for data labeling.

This work proposes two strategies to improve un-

supervised domain adaptation (UDA) in LiDAR se-

mantic segmentation, see a sample result on Fig. 1.

The ﬁrst strategy addresses this problem by apply-

ing a set of simple data processing steps to align the

data distribution reducing the domain gap on the in-

put space. The second strategy proposes how to align

the distribution on the output space by aligning the

class distribution. These two proposed strategies can

be applied in conjunction with current state-of-the-art

approaches boosting their performance.

We validate our approach on different scenar-

ios getting state-of-the-art results. We use the Se-

manticKitti (Behley et al., 2019) as the source domain

and we adapt it to SemanticPoss (Pan et al., 2020), to

Paris-Lille-3D (Roynard et al., 2018) and to a new

collected and released dataset.

2 RELATED WORK

This section describes the related work for the most

relevant related topics for this work: 3D LiDAR seg-

mentation and, unsupervised domain adaptation for

semantic segmentation.

2.1 3D LiDAR Point Cloud

Segmentation

3D LiDAR Semantic segmentation aims to assign a

semantic label to every point scanned by the LiDAR

sensor. Before the current trend and wide adoption of

deep learning approaches, earlier methods relied on

exploiting prior knowledge and geometric constraints

(Xie et al., 2019). As far as deep learning meth-

ods are concerned, there are two main types of ap-

proaches to tackle the 3D LiDAR semantic segmen-

tation problem. On one hand, there are approaches

that work directly on the 3D points, i.e., the raw point

cloud is taken as the input (Qi et al., 2017a; Qi et al.,

2017b). On the other hand, other approaches con-

vert this 3D point cloud into another representation

(images(Alonso et al., 2020), voxels(Zhou and Tuzel,

2018), lattices(Rosu et al., 2019)) in order to have a

structured input. For LiDAR semantic segmentation,

the most commonly used representation is the spher-

ical projection (Alonso et al., 2020; Milioto et al.,

2019; Wu et al., 2018). Several recent works (Mil-

ioto et al., 2019; Wu et al., 2018; Alonso et al., 2020)

show that point-based methods, i.e., approaches that

work directly on the 3D points, are slower and tend

to be less accurate than methods which project the 3D

point cloud into a 2D representation and make use of

convolutional layers.

2.2 Unsupervised Domain Adaptation

for Semantic Segmentation

Unsupervised Domain Adaptation (UDA) aims to

adapt models that have been trained on one speciﬁc

domain (source domain) to be able to work on a dif-

ferent domain (target domain) where there is a cer-

tain lack of labeled training data. Most works fol-

low similar ideas: the input data or features from

a source-domain sample and a target-domain sample

should be indistinguishable. Several works follow an

adversarial training scheme to minimize the distribu-

tion shift between the target and source domains data.

This approach has been shown to work properly at

pixel space (Yang and Soatto, 2020), at feature space

(Hoffman et al., 2018) and at output space (Vu et al.,

2019). However, adversarial training schemes, such

as DANN (Ganin et al., 2016), tend to present con-

vergence problems. Alternatively, other works follow

different schemes like CORL (Sun et al., 2017) or Mi-

nEnt (Vu et al., 2019). Entropy minimization meth-

ods (Vu et al., 2019; Chen et al., 2019) do not require

complex training schemes. They rely on a loss func-

tion that minimizes the entropy of the unlabeled target

domain output probabilities.

Domain Adaptation in LiDAR Semantic Segmentation by Aligning Class Distributions

331

Regarding segmentation on LiDAR data, very few

works have studied the problem of domain adaptation.

SqueezeSegV2 (Wu et al., 2019) proposes how to

adapt synthetic data where only coordinates are avail-

able without reﬂectance to real data. For this, they

propose a learned intensity rendering to create the in-

tensity for the synthetic data. A very recent work,

Xmuda (Jaritz et al., 2020) focuses on combining dif-

ferent modalities: LiDAR and RGB for multi-modal

domain adaptation. They propose to apply the KL

divergence between the output probabilities of both

modalities as the main loss function. Besides, they

also apply previously proposed methods like entropy

minimization (Vu et al., 2019). Differently, this work

investigates different UDA strategies (both existing

and novel) to improve UDA for the particular case of

urban LiDAR semantic segmentation. The presented

results show their effectiveness in reducing the do-

main gap.

3 UNSUPERVISED DOMAIN

ADAPTATION FOR LiDAR

SEMANTIC SEGMENTATION

This section describes the proposed domain adapta-

tion approach, including the LiDAR semantic seg-

mentation method used, the strategies proposed to re-

duce the domain gap (data alignment and class dis-

tribution alignment), and the formulation of the pro-

posed learning task. Figure 2 presents an overview of

our proposed approach which is further explained in

the following subsections.

3.1 LiDAR Semantic Segmentation

Model

We use a recent method for LiDAR semantic segmen-

tation which achieves state-of-the-art performance on

several LiDAR segmentation datasets, 3D-MiniNet

(Alonso et al., 2020). This method consists of three

main steps. First, it learns a 2D representation from

the 3D points. Then, this representation is fed to a 2D

Fully Convolutional Neural Network that produces a

2D semantic segmentation. These 2D semantic labels

are re-projected back to the 3D space and enhanced

through a post-processing module.

Let X

⊂ R

N×3

be a set of source domain LiDAR

point clouds along with associated semantic labels,

⊂ (1, C)

. Sample x

is a point cloud of size N

and y

(n,c)

provides the label of point (n) as one-hot

encoded vector. Let F be our LiDAR segmentation

network which takes a point cloud x

x and predicts a

probability distribution (size C classes) for each point

of the point-cloud F(x

x) = P

(n,c)

The parameters θ

of F are optimized to minimize

cross-entropy loss:

seg

, y

) = −

∑

n=1

∑

c=1

(n,c)

logP

(n,c)

, (1)

on source domain samples. Therefore, as the super-

vised semantic segmentation is concerned, the opti-

mization problem simply reads:

min

∑

∈X

seg

, y

). (2)

3.2 Data Alignment Strategies for

LiDAR

The problem of domain adaptation, i.e., data distribu-

tion misalignment, between X

and X

(a set of target

domain LiDAR point clouds), can be handled on the

network weights θ

but also modifying X

and X

order to align the distributions at the input space.

Next, we describe the different strategies for better

data alignment that we propose to improve LiDAR

domain adaptation.

XYZ-shift Augmentation. One of the main causes

of misalignment for LiDAR point clouds is the loca-

tion of the sensor. Since the point cloud values are

relative to the sensor origin, these changes cause vari-

ations affecting the whole point cloud. We perform

shifts up to ±2 meters on the Z-axis (height) and up

to ±5 meters on the Y-axis and X-axis.

Per-class Augmentation. We also propose to per-

form the augmentation independently per class, i.e.,

applying different data augmentation parameters for

every class. In particular, in this work, we perform

shifts up to ±1 meters on the Z-axis (height) and up

to ±3 meters on the Y-axis and X-axis.

Same Number of Beams. Besides the sensor place-

ment and orientation, a signiﬁcant difference between

sensors is the number of captured beams, which re-

sults in a more sparse or dense point cloud. We pro-

pose to match the data beams between the two do-

mains by reducing the data from the sensor with a

higher number of beams ending up with more homo-

geneous data within X

and X

Only Relative Features. Point-cloud segmentation

methods commonly use both absolute and relative

(between points) values of the input data. In order to

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

332

Figure 2: Approach overview. The ﬁgure shows our pipeline steps and optimization losses. First, we perform distribution

alignment on the input space, i.e., data alignment strategies. Then, we optimize the segmentation loss for source samples

where the labels are known and the class alignment and entropy losses for target data where no labels are available (See Sect.

3 for details). Green continuous arrows are used for target data and blue dotted arrows for source data.

be independent of absolute coordinates that are less

robust compared to relative coordinates, we propose

to use only relative features of the data. Therefore, we

propose to use only relative distances of every point

with respect to their neighbors.

3.3 Class Distribution Alignment

The domain shift appears due to many different fac-

tors. For example, different environments can present

quite different appearances, the spatial distribution of

objects may vary, the capturing set-up for different

scenarios can be totally different, etc. Depending on

the problem tackled and prior knowledge, we can hy-

pothesize which of these differences can be neglected

and assumed not to affect to the models we are learn-

ing. In this work, we tackle the domain adaptation

problem between different urban environments, since

all the datasets used are from urban scenarios. Tak-

ing this into account, although the data distribution

changes between the datasets, we can assume that the

class distribution is going to be very similar across

these scenarios since the distribution shift in urban

environments is mostly due to appearance and spatial

distribution changes. For example, we can assume

that if y

has a distribution of 90% road pixels and

10% car pixels, then y

will likely present a similar

distribution.

Our approach learns parameters θ

of F in such

a way that the predicted class distribution F(X

)

matches the real class distribution of Y

, i.e., the his-

togram representing the frequency of occurrence of

each class, previously computed in an ofﬂine fashion.

To do so, We propose to compute the KL-divergence

between these two class distribution as the class align-

ment loss

align

, Y

) =

∑

n=1

hist(Y

)log

hist(Y

)

(n)

. (3)

Therefore, the optimization problem reads as:

min

∑

∈X

align

, Y

). (4)

Equation 2 requires to compute the class distribu-

tion P

over the whole dataset. As this is computa-

tionally unfeasible, we compute it over the batch as

an approximation.

3.4 Optimization Formulation

The entropy loss is computed as in MinEnt (Vu et al.,

2019):

ent

) =

−1

log(C)

∑

n=1

∑

c=1

(n,c)

logP

(n,c)

, (5)

while the segmentation L

seg

and class alignment

align

losse are computed as detailed in previous sub-

sections.

During training, we jointly optimize the super-

vised segmentation loss L

seg

on source samples and

the class alignment loss L

align

and entropy loss L

ent

on target samples. The ﬁnal optimization problem is

formulated as follows:

Domain Adaptation in LiDAR Semantic Segmentation by Aligning Class Distributions

333

min

∑

seg

, y

) +

align

∑

align

, Y

)

ent

∑

ent

(6)

with λ

ent

and λ

align

as the weighting factors of the

alignment and entropy terms.

4 EXPERIMENTAL SETUP

This section details the setup used in our evaluation.

This includes the datasets used in the evaluation and

the training protocol we followed.

4.1 Datasets

We use four different datasets for the evaluation.

They were collected in four different geographical

areas, with four different LiDAR sensors, and with

four different set-ups. We take the well known Se-

manticKITTI dataset (Behley et al., 2019) as the

source domain dataset and the other three datasets as

target domain data.

SemanticKITTI. The SemanticKITTI dataset

(Behley et al., 2019) is a recent large-scale dataset

that provides dense point-wise annotations. The

dataset consists of over 43000 LiDAR scans from

which over 21000 are available for training. The

dataset distinguishes 22 different semantic classes.

The capturing sensor is a Velodyne HDL-64E

mounted on a car.

Paris-Lille-3D. Paris-Lille-3D (Roynard et al.,

2018) is a medium-size dataset that provides three

aggregated point clouds. It is collected with a tilted

rear-mounted Velodyne HDL-32E placed on a vehi-

cle. Following PolarNet work (Zhang et al., 2020),

we extract individual scans from the registered point

clouds thanks to the scanner trajectory and points’

timestamps. Each scan is made of points within +/-

100m. We take the Lille-1 point cloud for the do-

main adaptation and Lille-2 for validation. We use

the following intersecting semantic classes with the

SemanticKitti: car, person, road, sidewalk, building,

vegetation, pole, and trafﬁc light.

SemanticPoss. The SemanticPoss (Pan et al., 2020)

is a medium-size dataset which contains 5 different

sequences from urban scenarios providing 3000 Li-

DAR scans. The sensor used is a 40-line Pandora

mounted on a vehicle. We take the three ﬁrst se-

quences for applying the adaptation methods and the

last two sequences for validation. We use the follow-

ing intersecting classes with the SemanticKitti: car,

person, trunk, vegetation, trafﬁc sign, pole, fence,

building, rider, bike, and ground (combines road and

sidewalk).

I3A. We have captured a small dataset to test our ap-

proach in a different scenario. In contrast to the three

previous datasets, this dataset is not captured from a

vehicle but from a TurtleBot, having the sensor at a

lower height than in the other set-ups. The captur-

ing sensor is the Velodyne VLP-16. We capture an

urban environment similar to previous datasets. The

dataset contains two sequences, one for training and

another for validation. We use the intersecting seman-

tic classes with the SemanticKitti: car, person, road,

sidewalk, building, vegetation, trunk, pole, and trafﬁc

light.

4.2 Training Protocol

As we mentioned in Sec. 3.1, we use 3D-MiniNet

(Alonso et al., 2020) as the base LiDAR semantic

segmentation method. In particular, we use the avail-

able 3D-MiniNet-small version because of memory

issues. For computing the relative coordinates and

features, we follow 3D-MiniNet approach extracting

them from the N neighbors of each 3D point where N

is set to 16.

For all the experiments we train this architecture

for 700K iterations with a batch size of 8. We use

Stochastic Gradient Descent (SGD) optimizer with an

initial learning rate of 0.005 and a polynomial learn-

ing rate decay schedule with a power set to 0.9.

We set λ

ent

to 0.001 as suggested in MinEnt (Vu

et al., 2019) and λ

align

to 0.001. We empirically no-

ticed that the performance is very similar when these

two hyper-parameters are set between 10

−5

and 10

−2

The two main conditions for them to work properly

are: (1) be greater than 0 and, (2) do not be higher

than the supervised loss.

One thing to take into account is that, as explained

in 4.1, the Paris-Lille-3D has a very limited ﬁeld of

view. Therefore in order to make MinEnt (Vu et al.,

2019) work in this dataset, we had to simulate the

same ﬁeld of view on the source dataset.

5 RESULTS

This section presents the experimental validation of

our approach compared to different baselines. The

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

334

Table 1: Ablation study of our domain adaption pipeline

for semantic segmentation. Source dataset: SemanticKitti

(Behley et al., 2019).

mIoU on mIoU on mIoU on

Target dataset I3A ParisLille SemanticPoss

Base model 15.9 19.2 14.0

+ XYZ-shift augmentation 25.1 28.9 16.8

+ Per-class augmentation 27.0 30.1 18.2

+ Same number of beams 42.0 35.4 19.7

+ Only relative features 47.1 — 21.8

+ MinEnt (Vu et al., 2019) 50.3 41.5 26.2

+ Class distrib. alignment 52.5 42.7 27.0

— Not used because there was no performance gain.

proposed approach achieves better results than the

other baselines in the three different scenarios for

unsupervised domain adaptation in LiDAR Seman-

tic Segmentation. In all the experiments we use the

SemanticKITTI dataset (Behley et al., 2019) as the

source data distribution and perform the adaptation on

the other three datasets.

5.1 Ablation Study

The experiments in this subsection show how the dif-

ferent data alignment steps and the proposed learning

losses affect the ﬁnal performance of our approach.

Table 1 summarizes the ablation study performed on

three different scenarios. The results show how all

the steps proposed contribute towards the ﬁnal per-

formance. The main insights observed in the ablation

study are discussed next.

Performing strong XYZ-shifts results in a boost on

the performance, meaning that the domain gap is con-

siderably reduced. The distribution gap reduced by

this step is the one caused by the fact of using different

LiDAR sensor set-ups (such as different acquisition

sensor height). Besides, in these autonomous driving

set-ups, the distance between the car and the objects

depends on how wide are the streets or on which lane

is the data capturing source. Therefore, this is an es-

sential and really easy data transformation to perform

which gives an average of 7.2% MIoU gain.

The per-class data augmentation also boosts the

performance. This data augmentation method tries

to reduce the domain gap by adding different relative

distances between different classes gaining an aver-

age of 1.3% MIoU gain.

Another interesting and straightforward technique

to perform is to match the number of LiDAR beams of

the source and target data, i.e., match LiDAR point-

cloud resolution. This helps the data alignment es-

pecially for having the same point density on the

3D point-cloud and similar relative distances between

the points. We show that this method gives an im-

provement similar to the XYZ-shift data augmenta-

tion, hugely reducing the domain gap. The higher the

initial difference in the number of beams, the more

improvement we can get: the i3A LiDAR has 16

beams, the ParisLille 32, and the SemanticPoss 40,

compared to the 64 of the source data (SemanticKitti).

The use of relative features only does not always

help to reduce the domain gap, it was only beneﬁ-

cial on the i3A and SemanticPoss datasets. Remov-

ing the absolute features and only learning from rel-

ative features helps especially when the relative dis-

tances between the 3D points have less domain shift

than the absolute coordinates. This will depend on the

dataset, but the stronger the differences between cap-

turing sensors, the more likely that the use of relative

features will help.

Besides the data alignment steps, our approach in-

cludes two learning losses to the pipeline to help to re-

duce the domain gap. The ﬁrst one is the entropy min-

imization loss proposed in MinEnt (Vu et al., 2019).

The second one is our proposed class distribution

alignment loss introduced in this work. We show

that these two losses can be combined for the do-

main adaptation problem and that, although less sig-

niﬁcantly with respect to previously discussed steps,

they also improve on the three different set-ups, con-

tributing to achieving state-of-the-art performance.

5.2 Comparison with Other Baselines

Table 2: Results on the three different LiDAR semantic seg-

mentation datasets using different domain adaptation meth-

ods. The source dataset is the SemanticKitti dataset (Behley

et al., 2019).

mIoU on mIoU on mIoU on

I3A ParisLille SemanticPoss

Baseline 15.9 19.2 13.4

MinEnt (Vu et al., 2019) 28.4 23.2 19.6

AdvEnt (Vu et al., 2019) 21.0 20.7 19.5

MaxSquare (Chen et al., 2019) 28.4 22.8 19.3

Data alignment (ours)* 47.1 36.2 19.0

Full approach (ours) 52.5 42.7 27.0

* Only data alignment strategies from Sect. 3.2

Table 2 shows the comparison of our pipeline (com-

posed of all the steps discussed in the ablation study)

with other existing methods for domain adaptation.

We select MinEnt, Advent, and MaxSquare as the

baselines because they are leading the state-of-the-

art for unsupervised domain adaptation. We use the

available authors’ code for replication.

We apply the different domain adaptation meth-

ods of the three different set-ups without our data

alignment steps. This comparison shows that good

pre-processing of the data can obtain better results

than just applying out-of-the-box methods for domain

adaptation. It also shows that our complete pipeline

outperforms these previous methods on LiDAR do-

main adaptation. Our results demonstrate that com-

Domain Adaptation in LiDAR Semantic Segmentation by Aligning Class Distributions

335

Figure 3: LiDAR domain adaptation results for different methods and datasets: I3A dataset ﬁrst row, ParisLille dataset second

row, and SemanticPoss last row. From left to right: Input point cloud, ground truth labels, baseline with no adaptation (trained

on SemanticKitti), MinEnt (Vu et al., 2019), our adaptation only with data processing strategies, our full adaptation pipeline.

Best viewed in color. For better and more detail (video result) go to https://youtu.be/EpkJ UH1F-o.

bining proper data processing with learning methods

for domain adaptation gives an average of more than

×2 boost on the performance.

Figure 3 includes a few examples of the seg-

mentation obtained with a baseline with no domain

adaptation, using the MinEnt (Vu et al., 2019) ap-

proach only, with our approach using only the data

pre-processing steps, and with our approach includ-

ing all steps proposed. We can appreciate in ﬁg-

ure 3 how data processing helps on certain semantic

classes, such as road, person, car, or vegetation, while

MinEnt usually improves at different ones like build-

ing. This suggests the good complementary of both

strategies, and indeed combining them provides the

best results. More detail in additional video results

can be seen at https://youtu.be/EpkJ UH1F-o.

6 CONCLUSIONS

In this work, we introduce a novel pipeline that ad-

dresses the task of unsupervised domain adaptation

for LiDAR semantic segmentation. Our pipeline con-

sists of aligning data distributions on the data space

with different simple strategies combined with learn-

ing losses on the semantic segmentation process that

also force the data distribution alignment. Our re-

sults show that a proper data alignment on the input

space can produce better domain adaptation results

that just using out-of-the-box state-of-the-art learn-

ing methods. Besides, we show that combining these

data alignment methods with learning methods, like

the one proposed in this work to align the class distri-

butions of the data, can reduce even more the domain

gap getting new state-of-the-art results. Our approach

is validated on three different scenarios, from differ-

ent datasets, as the target domain, where we show

that our full pipeline improves previous methods on

all three scenarios.

ACKNOWLEDGMENTS

This project was partially funded by projects FEDER/

Ministerio de Ciencia, Innovaci

on y Universidades/

Agencia Estatal de Investigaci

on/RTC-2017-6421-7,

PGC2018-098817-A-I00 and PID2019-105390RB-

I00, Arag

on regional government (DGA T45

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

336

17R/FSE) and the Ofﬁce of Naval Research Global

project ONRG-NICOP-N62909-19-1-2027.

REFERENCES

Alonso, I., Riazuelo, L., Montesano, L., and Murillo, A. C.

(2020). 3d-mininet: Learning a 2d representation

from point clouds for fast and efﬁcient 3d lidar se-

mantic segmentation. IEEE Robotics and Automation

Society, RAL.

Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke,

S., Stachniss, C., and Gall, J. (2019). SemanticKITTI:

A dataset for semantic scene understanding of lidar

sequences. In Proceedings of the IEEE International

Conference on Computer Vision.

Chen, M., Xue, H., and Cai, D. (2019). Domain adaptation

for semantic segmentation with maximum squares

loss. In Proceedings of the IEEE International Con-

ference on Computer Vision, pages 2090–2099.

Ganin, Y., Ustinova, E., Ajakan, H., Germain, P.,

Larochelle, H., Laviolette, F., Marchand, M., and

Lempitsky, V. (2016). Domain-adversarial training of

neural networks. The journal of machine learning re-

search, 17(1):2096–2030.

Hoffman, J., Tzeng, E., Park, T., Zhu, J.-Y., Isola, P.,

Saenko, K., Efros, A., and Darrell, T. (2018). Cycada:

Cycle-consistent adversarial domain adaptation. In

International conference on machine learning, pages

1989–1998.

Jaritz, M., Vu, T.-H., Charette, R. d., Wirbel, E., and P

erez,

P. (2020). xmuda: Cross-modal unsupervised domain

adaptation for 3d semantic segmentation. In Proceed-

ings of the IEEE/CVF Conference on Computer Vision

and Pattern Recognition, pages 12605–12614.

Li, Y., Yuan, L., and Vasconcelos, N. (2019). Bidirectional

learning for domain adaptation of semantic segmenta-

tion. In Proceedings of the IEEE Conference on Com-

puter Vision and Pattern Recognition, pages 6936–

6945.

Mei, J., Gao, B., Xu, D., Yao, W., Zhao, X., and Zhao,

H. (2019). Semantic segmentation of 3d lidar data in

dynamic scene using semi-supervised learning. IEEE

Transactions on Intelligent Transportation Systems,

21(6):2496–2509.

Milioto, A., Vizzo, I., Behley, J., and Stachniss, C. (2019).

Rangenet++: Fast and accurate lidar semantic seg-

mentation. In Proceedings of the IEEE/RSJ Interna-

tional Conference on Intelligent Robots and Systems

(IROS).

Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., and Zhao, H.

(2020). Semanticposs: A point cloud dataset with

large quantity of dynamic instances. arXiv preprint

arXiv:2002.09147.

Qi, C. R., Su, H., Mo, K., and Guibas, L. J. (2017a). Point-

net: Deep learning on point sets for 3d classiﬁcation

and segmentation. In Proceedings of the IEEE Con-

ference on Computer Vision and Pattern Recognition,

pages 652–660.

Qi, C. R., Yi, L., Su, H., and Guibas, L. J. (2017b). Point-

net++: Deep hierarchical feature learning on point sets

in a metric space. In Advances in neural information

processing systems, pages 5099–5108.

Rosu, R. A., Sch

utt, P., Quenzel, J., and Behnke, S. (2019).

Latticenet: Fast point cloud segmentation using per-

mutohedral lattices. arXiv preprint arXiv:1912.05905.

Roynard, X., Deschaud, J.-E., and Goulette, F. (2018).

Paris-lille-3d: A large and high-quality ground-truth

urban point cloud dataset for automatic segmenta-

tion and classiﬁcation. The International Journal of

Robotics Research, 37(6):545–557.

Sun, B., Feng, J., and Saenko, K. (2017). Correlation align-

ment for unsupervised domain adaptation. In Domain

Adaptation in Computer Vision Applications, pages

153–171. Springer.

Tzeng, E., Hoffman, J., Saenko, K., and Darrell, T. (2017).

Adversarial discriminative domain adaptation. In Pro-

ceedings of the IEEE conference on computer vision

and pattern recognition, pages 7167–7176.

Vu, T.-H., Jain, H., Bucher, M., Cord, M., and P

erez, P.

(2019). Advent: Adversarial entropy minimization for

domain adaptation in semantic segmentation. In Pro-

ceedings of the IEEE conference on computer vision

and pattern recognition, pages 2517–2526.

Wang, M. and Deng, W. (2018). Deep visual domain adap-

tation: A survey. Neurocomputing, 312:135–153.

Wu, B., Wan, A., Yue, X., and Keutzer, K. (2018). Squeeze-

seg: Convolutional neural nets with recurrent crf for

real-time road-object segmentation from 3d lidar point

cloud. In 2018 IEEE International Conference on

Robotics and Automation (ICRA), pages 1887–1893.

IEEE.

Wu, B., Zhou, X., Zhao, S., Yue, X., and Keutzer, K.

(2019). Squeezesegv2: Improved model structure

and unsupervised domain adaptation for road-object

segmentation from a lidar point cloud. In 2019 In-

ternational Conference on Robotics and Automation

(ICRA), pages 4376–4382. IEEE.

Xie, Y., Tian, J., and Zhu, X. X. (2019). A review of

point cloud semantic segmentation. arXiv preprint

arXiv:1908.08854.

Yang, Y. and Soatto, S. (2020). Fda: Fourier domain adap-

tation for semantic segmentation. In Proceedings of

the IEEE/CVF Conference on Computer Vision and

Pattern Recognition, pages 4085–4095.

Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B.,

and Foroosh, H. (2020). Polarnet: An improved grid

representation for online lidar point clouds semantic

segmentation. In Proceedings of the IEEE/CVF Con-

ference on Computer Vision and Pattern Recognition,

pages 9601–9610.

Zhou, Y. and Tuzel, O. (2018). Voxelnet: End-to-end learn-

ing for point cloud based 3d object detection. In Pro-

ceedings of the IEEE Conference on Computer Vision

and Pattern Recognition, pages 4490–4499.

Zou, Y., Yu, Z., Liu, X., Kumar, B., and Wang, J. (2019).

Conﬁdence regularized self-training. In Proceedings

of the IEEE International Conference on Computer

Vision, pages 5982–5991.

Domain Adaptation in LiDAR Semantic Segmentation by Aligning Class Distributions

337