Segmentation of Fish in Realistic Underwater Scenes using Lightweight

Deep Learning Models

Gordon B

oer

, Rajesh Veeramalli

and Hauke Schramm

1,2

Institute of Applied Computer Science, Kiel University of Applied Sciences, Kiel, Germany

Department of Computer Science, Faculty of Engineering, Kiel University, Germany

Keywords:

Semantic Segmentation, Underwater Imagery, Fish Detection.

Abstract:

The semantic segmentation of ﬁsh in real underwater scenes is a challenging task and an important prerequisite

for various processing steps. With a good segmentation result, it becomes possible to automatically extract

the ﬁsh contour and derive morphological features, both of which can be used for species identiﬁcation and

ﬁsh biomass assessment. In this work, two deep learning models, DeepLabV3 and PSPNet, are investigated

for their applicability to ﬁsh segmentation for a ﬁsh stock monitoring application with low light cameras. By

pruning these networks and employing a different encoder, they become more suitable for systems with lim-

ited hardware, such as remotely operated or autonomously operated underwater vehicles. Both segmentation

models are trained and evaluated on a novel dataset of underwater images showing Gadus morhua in its natural

behavior. On a challenging test set, which includes ﬁsh recorded at difﬁcult visibility conditions, the PSPNet

performs best, and achieves an average pixel accuracy of 96.8% and an intersection-over-union between the

predicted and the target mask of 73.8%. It achieves this with a very limited parameter set of 94,393 trainable

parameters.

1 INTRODUCTION

Digital imaging in marine research has become a stan-

dard tool to help marine biologists answer many sci-

entiﬁc questions. This is largely due to rapid techno-

logical advances in recent decades that have resulted

in digital cameras with higher technological capabil-

ities, such as better image quality, larger storage ca-

pacities, and better in-situ applicability, while at the

same time being available at lower prices in the con-

sumer market. The applications of modern sensors

and algorithms for underwater imaging are numerous

and several reviews have been published, emphasiz-

ing either their general applicability for marine sci-

ence (Durden et al., 2016; Malde et al., 2020; Fer-

nandes et al., 2020) or for more speciﬁc research ar-

eas, like the observation of coastal marine biodiver-

sity (Mallet and Pelletier, 2014), the monitoring of hu-

man impact on marine environments (Bicknell et al.,

2016), the automatic determination of ﬁsh species

(Alsmadi and Almarashdeh, 2020) or the investiga-

tion of ﬁsh connectivity (Lopez-Marcano et al., 2021).

The aim of this work is to investigate the applica-

bility of deep learning methods to segment and outline

ﬁshes, detected in real-world underwater scenes. In

addition to the information about what kind of objects

are present in an image and where they are located, a

successful semantic segmentation reveals what class

each pixel belongs to. It thereby becomes possible, to

additionally extract the outline of an object of inter-

est and the concise area it covers in the image. The

precise segmentation of a ﬁsh is an important pre-

requisite for the automatic determination of morpho-

metric characteristics, like the total length, which in

turn can be used to determine the ﬁsh weight (Wil-

helms et al., 2013). Additionally, well-deﬁned land-

marks on the ﬁsh outline are commonly used to iden-

tify speciﬁc ﬁsh species (Rawat et al., 2017; Cav-

alcanti et al., 1999), while automatically localizing

those landmarks becomes easier with a good segmen-

tation result. The presented algorithm is an impor-

tant software part for a currently developed underwa-

ter sensor platform, that aims to estimate ﬁsh biomass

for a ﬁsh stock assessment, using non-invasive sensor

technology.

Recently, there has also been a great demand to

apply the undoubtedly successful deep learning al-

gorithms using limited hardware. This is especially

true for remotely controlled and autonomous vehicles,

such as underwater robots, for which we intend to use

158

Böer, G., Veeramalli, R. and Schramm, H.

Segmentation of Fish in Realistic Underwater Scenes using Lightweight Deep Learning Models.

DOI: 10.5220/0010712700003061

In Proceedings of the 2nd International Conference on Robotics, Computer Vision and Intelligent Systems (ROBOVIS 2021), pages 158-164

ISBN: 978-989-758-537-1

the studied algorithms in the future. Therefore, more

light-weight segmentation models are investigated in

this publication. Precisely speaking, pruned versions

of DeepLabv3 (Chen et al., 2017) and PSPNet (Zhao

et al., 2017) are used on image extracts containing ﬁsh

to perform a binary segmentation into a ﬁsh and back-

ground class.

All the used underwater videos were recorded

with speciﬁc low-light cameras adapted for an under-

water usage, that allowed the recording of underwater

scenes without active lighting.

The main contributions of this work are:

• Alternative conﬁgurations of PSPNet and

DeepLabv3 which are better suited for hardware-

limited devices.

• The comparison of those segmentation models re-

garding their inference times and performance for

ﬁsh segmentation.

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

tion 2 provides an overview of related publications. In

section 3, the used recording setup and the employed

algorithms, together with the metrics used to evaluate

them, are described. Section 4 details the experimen-

tal evaluation, and section 5 concludes this paper.

2 RELATED WORK

Besides the automatic localization and classiﬁcation

of arbitrary objects in digital images, the semantic

segmentation of those images is another active ﬁeld

of research. It is of great relevance in many applica-

tion areas like autonomous driving (Grigorescu et al.,

2020), remote sensing (Marmanis et al., 2016), 3D-

sensing (Tchapmi et al., 2017) or cancer prediction

(Kourou et al., 2015). As in many areas of com-

puter vision, the most powerful algorithms to date

are based on deep learning architectures, well es-

tablished ones being Fully Convolutional Networks

(Long et al., 2015), Mask R-CNN (He et al., 2017), U-

Net (Ronneberger et al., 2015), PSPNet (Zhao et al.,

2017) or SegNet (Badrinarayanan et al., 2017).

Regarding the segmentation of ﬁsh in digital im-

ages, several efforts have been published so far. We

will focus on methods that have been tested on free-

swimming ﬁsh, recorded in underwater scenarios.

Due to the different settings, the results obtained for

dead ﬁsh photographed in air, e.g. on a photo-table

(Yu et al., 2020; Konovalov et al., 2019; Baloch et al.,

2017), conveyor belt (Storbeck and Daan, 2001) or

ﬁsh-trawler (Yang et al., 2018; French et al., 2015),

are not directly comparable.

So far, various standard algorithms have been ap-

plied to extract segmentation masks for ﬁsh, like

Otsu thresholding, edge detection, Grabcut, mean-

shift, matrix decomposition or the curvature scale

space transform (Abdeldaim et al., 2018; Qin et al.,

2014; Spampinato et al., 2010; Abbasi and Mokhtar-

ian, 1999). However, all those algorithms do not

proﬁt from recent advances possible with novel deep

learning methods, e.g. the possibility to automatically

learn important image features, as opposed to hand-

crafted ones.

To our knowledge, published efforts to use deep

learning methods for semantic segmentation of free-

swimming marine animals are still rather limited.

Recently, the Enhanced Cascade Decoder Network

(ECD-Net) has been proposed (Li et al., 2021), which

is utilized for the segmentation of marine animals,

including several ﬁsh species. It builds upon a pre-

trained ResNet-50 (He et al., 2016) backbone, fol-

lowed by several feature enhancement and cascade

decoder modules which are trained using a mixture

loss. The authors report superior results, on a self-

published dataset, as compared to several state-of-the-

art models. Although the ECD-Net, in terms of train-

able parameters, has been reported to be smaller than

most of the other considered networks, with the re-

maining 207 million parameters it is much too com-

plex for the usage in an embedded system. In another

work, images from the stereoscopic camera system

Deep Vision (Rosen and Holst, 2013), which records

the water volume at the opening of a net trawl, were

automatically processed to determine the length of

visible ﬁsh. For this purpose, a Mask R-CNN was

used to detect and segment ﬁsh, followed by a re-

ﬁnement step to distinguish between individual over-

lapping ﬁsh. The authors report an average IoU of

84.5% for an independent test set of 200 images (Gar-

cia et al., 2020). Given the used camera setup, the ﬁsh

is recorded showing an unnatural behavior in front of

a rather uniform background, therefore, the obtained

results may not be directly transferable to real world

underwater scenarios.

3 MATERIALS AND METHODS

3.1 Data Acquisition

All experiments were carried out on videos of realis-

tic underwater scenes, which are part of a larger col-

lection of underwater sensor data, recorded by a pro-

totype underwater sensor platform. The device was

developed in a joint project of German universities

and marine engineering companies. Over the period

Segmentation of Fish in Realistic Underwater Scenes using Lightweight Deep Learning Models

159

of several months, the sensor platform was deployed

at seaﬂoor level, at a depth of 22m in the North Sea,

about 45 nautical miles west of the island of Sylt, and

in the Kiel Fjord, an inlet of the Baltic Sea in north-

ern Germany. During the measurement campaigns,

the primary goal was to conduct a continuous record-

ing of stereo video and sonar data as well as various

oceanic parameters at a high temporal sampling rate.

In total, the raw data set spans approximately 240

days, resulting in nearly 3000 hours of video footage,

as the cameras were not operated during the night

hours.

All stereo videos were recorded using two Photo-

nis ”Nocturn XL” monochrome CMOS image sensors

with an optical resolution of 1280x1024 pixels and a

sampling rate of up to 20 frames/second, housed in

a specially designed underwater case with a ﬂat view

port. The selected camera system is particularly suit-

able for low-light scenarios, which made it possible

to record without active lighting from dawn to dusk,

which was veriﬁed to a maximum depth of 22m. The

exclusive use of passive lighting is justiﬁed by the

main concept of the platform, which is to be as less

invasive as possible, e.g. by avoiding attraction ef-

fects from light sources (Marchesan et al., 2005), thus

ensuring that the ﬁsh can be recorded in their natural

behavior.

The same stereo camera setup will be duplicated

on a mobile platform that is currently being built,

making it possible to reuse many of the insights

gained from the stationary system as has been used

in the present work.

3.2 Methodology

In general, the semantic segmentation of an image can

be deﬁned as a pixel-wise classiﬁcation, where each

pixel is assigned to a speciﬁc object class. Depending

on the complexity, the task can further be divided into:

• Binary segmentation, if only 2 classes are sepa-

rated, e.g. foreground and background.

• Multi-class segmentation, if multiple classes are

considered, e.g. car, pedestrian, street, building.

• Instance segmentation, if each pixel additionally

is assigned to a unique instance of an object class,

thereby making it possible to distinguish between

several, possibly overlapping, occurrences of the

same object class.

The problem addressed in this work is a binary seg-

mentation problem since only the two classes ﬁsh and

background are considered.

A simple approach to deﬁne a segmentation net-

work is to stack multiple convolutional layers with the

same padding to preserve the resolution of the input

image. This type of architecture is computationally

intensive, as it preserves the input resolution through

all layers of the network. Therefore, most recent seg-

mentation networks follow an encoder-decoder archi-

tecture, which is comparatively more efﬁcient. First,

a sequence of convolutional and either downsampling

or pooling layers creates a low-resolution image rep-

resentation which encodes the high-resolution infor-

mation of the input image. Second, to reconstruct the

original resolution in the output image, again a se-

quence of convolutional layers, followed by upsam-

pling layers or transposed convolutions is added, to

gradually increase the size of the spatial features.

In the present work, the two segmentation models

DeepLabv3 and PSPNet are used, both of which fol-

low the described type of encoder-decoder architec-

ture. As opposed to the original architectures, which

utilize a ResNet (He et al., 2016) as decoder, we em-

ploy MobileNetV2 (Sandler et al., 2018) for each of

the two segmentation networks. MobileNetV2 makes

use of depthwise convolutions and inverted residuals,

which are grouped together into several subsequent

blocks using ReLU6 as non-linearity function. The

features from the ﬁrst 3 stages of the MobileNetV2,

which has been pre-trained on ImageNet, are passed

to the decoder part of the respective network. These

changes reduce the model complexity and size, which

makes them better suited for systems with limited

hardware.

3.2.1 PSPNet

In the original implementation of PSPNet, a pre-

trained ResNet was employed in combination with a

dilated network strategy to extract the feature maps.

As was mentioned above, the ResNet was inter-

changed with MobileNetV2 in the proposed system.

The spatial pyramid pooling (SPP) module in PSP-

Net ﬁrst downsamples the feature maps from the en-

coder at 4 different scales (1 x 1, 2 x 2, 3 x 3 and 6

x 6), all of which are afterwards upsampled and fused

together. The integrated feature maps from the SPP

module are concatenated along with the feature maps

from the encoder. The decoder, followed by a bilinear

upsampling layer with a scale of 8, then converts the

concatenated feature maps to the segmentation out-

put. The SPP module eliminates the requirement for

a ﬁxed input size.

3.2.2 DeepLabv3

In the DeepLabv3 architecture, a series of 3 x 3 atrous

convolutions are built in cascade, which are able

to capture long-range information from the inputs.

ROBOVIS 2021 - 2nd International Conference on Robotics, Computer Vision and Intelligent Systems

160

Since in the original architecture, the last block of the

ResNet encoder is duplicated 3 times, we are dupli-

cating the last blocks of the MobileNetV2 encoder as

well. The atrous spatial pyramid pooling module in

DeepLabv3 consists of multiple parallel atrous con-

volutional layers with different dilation rates. This

module resamples the feature maps at multiple rates,

which have a complementary ﬁeld of view compared

to normal convolution ﬁlters. Following this idea, it

becomes possible to classify objects or regions of ar-

bitrary size.

3.3 Evaluation Metrics

The metrics used for the validation of the models are

the pixel-level accuracy and Intersection-over-Union

(IoU):

• The overall per-pixel accuracy measures the ratio

of correctly classiﬁed pixels to total pixels.

• The IoU is measured as the area of overlap, di-

vided by the area of union, between the predicted

segmentation and the ground truth segmentation.

This metric is often referred to as Jaccard Index.

The investigated task of ﬁsh segmentation can be de-

ﬁned as an unbalanced, binary segmentation problem,

since the background class dominates the ﬁsh class

in the dataset. In this case, the pixel-level accuracy

is sensitive to the class-imbalance problem, which is

why we rely on the IoU as an additional metric.

4 EXPERIMENTAL RESULTS

In this section, we provide a brief description of the

used data, the implementation and training details,

and the results achieved, as compared for the two net-

works.

4.1 Data

The used dataset consists of 600 labeled grayscale im-

ages with a resolution of 1280x1024 pixels, that de-

pict ﬁsh of the species Gadus morhua, which is of sig-

niﬁcant importance to the ﬁshery industry. The full

image set was randomly split into 80% for training,

10% for testing and 10% for validating the models.

Since the images in these data splits have been ex-

tracted from video sequences, very similar samples

may appear in the test and training set due to the

simple random selection, e.g. if a ﬁsh is swimming

slowly in front of the camera. Because of this, a sep-

arate test set of 1148 images, that where recorded on

a different day have been annotated as well. Using

this test set, we aim to fairly assess the generalization

ability of the segmentation models used. All images

have been annotated with binary segmentation masks,

where pixels, belonging to a ﬁsh, are marked as fore-

ground, everything else as background, respectively.

In most of the samples, background pixels predom-

inate the number of foreground pixels. Each thereby

annotated ﬁsh is included in a bounding-box of a ﬁxed

size of 512x512 pixels. In several cases, a bounding

box may contain more than one ﬁsh.

Figure 1: Examples of an annotated segmentation mask.

4.2 Training Setup

The models have been implemented with PyTorch and

were trained and evaluated on a NVIDIA TITAN XP

GeForce RTX 2080 TI GPU. For the model optimiza-

tion, the Dice loss was used to determine the error

between the prediction and the ground truth, which is

calculated at pixel-level. The Dice loss for a complete

image is deﬁned by the formula:

DL =

∑

ˆp

∑

ˆp

∑

(1)

with y

and ˆp

being the values of corresponding pix-

els representing the true object class and the predicted

class for this pixel, respectively. The network param-

eters were optimized using the Adam backpropaga-

tion algorithm, with a learning rate of 0.001. Both

models were trained with a batch size of 4, each sam-

ple being ﬂipped with a probability of 0.5 in the hori-

zontal or vertical direction. Other augmentation tech-

niques were not used, since the possible data inter-

polation may produce an unnatural ﬁsh appearance.

Both models were trained for 30 epochs on the same

data splits. All segmentation results were thresholded

at a conﬁdence value of 0.5 to obtain binary masks.

Segmentation of Fish in Realistic Underwater Scenes using Lightweight Deep Learning Models

161

4.3 Results and Discussion

The performance of each model on the independent

test set of 1148 images, using the aforementioned

evaluation metrics, is listed in Table 1. Regarding the

pixel accuracy, both models perform almost equally

well with an average accuracy of 96.8% for PSPNet

and 96.2% for DeepLabv3, respectively. Consider-

ing the IoU, PSPNet performs better with an average

value of 73.8% compared to the 69.9% as achieved by

DeepLabv3.

Table 1: Average pixel level accuracy and IoU as obtained

by the two utilized models.

Model Pixel Accuracy IoU

PSPNet 96.8% 73.8%

DeepLabv3 96.2% 69.9%

To assess which model is better suited for con-

strained hardware devices, we investigate their re-

spective inference times and model sizes, as listed

in Table 2. While both models perform equally well

in terms of segmentation accuracy, it is obvious, that

the PSPNet achieves this by using much less mem-

ory at a slightly higher framerate. The difference in

model size can be explained by the number of train-

able parameters, which adds up to 94,393 for the

PSPNet and 121,382,5 for DeepLabv3, respectively.

Deeplabv3 has much more trainable parameters be-

cause of the used trainable atrous convolutions in the

decoder. Given this insight, we suggest using the

PSPNet for a hardware limited device.

Table 2: Inference speed, in frames per second (FPS), and

model size of the two adapted models.

Model

Inference speed (FPS)

Model size

CPU GPU

PSPNet 5.5 155 463 KB

DeepLabv3 3.45 148 5 MB

We have investigated those examples, where both

models fail to generate a perfect segmentation mask,

which revealed, that both models made similar errors

on the same samples. As is illustrated in Figure 2,

smaller errors can occur in the border regions of a ﬁsh.

Although the larger part of the ﬁsh body and the back-

ground have been segmented correctly, a small dis-

crepancy can be observed at the very edge of the ﬁsh.

In our opinion, this happens largely due to an inaccu-

rate ground-truth mask, which was annotated rather

coarsely, while the automatically generated mask is

characterized by much smoother and more detailed

edges. Another source of error are examples with

a low signal-to-noise ratio as depicted in Figure 3.

Those are typically images which were recorded at

Figure 2: Illustration of a good segmentation result ob-

tained with PSPNet, showing the input image, the predicted

segmentation, the ground-truth and the difference between

them. The difference between the predicted and the target

mask is highlighted by white pixels.

Figure 3: Example of a partially false segmentation, due to

difﬁcult viewing conditions.

times of twilight, i.e. when the sun was below the

horizon. A comparable case is shown in Figure 4,

in which the ﬁsh becomes hardly visible while swim-

ming out of the imaged area, towards the ocean ﬂoor.

Since the recording setup does not use active lighting,

to avoid an unnatural attraction effect, the lower part

of the image is less well illuminated by the passive

lighting from the surface. This explains the poor visi-

bility in the lower part of the image. For those cases,

the models were not able to provide a good segmenta-

tion, although the human annotator was able to clearly

mark the ﬁsh. To understand this effect, it is impor-

tant to consider that the human annotator could switch

between frames during the annotation process, reveal-

ing the movement of the ﬁsh and making it easier for

the human eye to see or guess even barely visible out-

lines. The segmentation models, on the other hand,

only work on single frames, without any knowledge

of the temporal context. Therefore, a possible exten-

ROBOVIS 2021 - 2nd International Conference on Robotics, Computer Vision and Intelligent Systems

162

sion of the algorithm would be to operate on the full

video sequence, utilizing the connected frames, e.g.

by an object tracking or averaging of generated seg-

mentation masks.

Figure 4: Example of a partially false segmentation, due to

the ﬁsh swimming out of the imaged area.

5 CONCLUSIONS AND FUTURE

WORK

In this paper, we have investigated PSPNet and

DeepLabv3 for their applicability to ﬁsh segmenta-

tion for a ﬁsh stock monitoring application with low

light cameras. The original architectures of both mod-

els were adapted for a usage on limited hardware,

by interchanging their respective decoder with Mo-

bileNetV2 and by reducing the number of used layers

in the encoder. We have trained and evaluated the seg-

mentation and processing performance of each model

on a custom dataset, which depicts freely swimming

ﬁshes in an unconstrained underwater environment. A

larger test set of 1148 images was used to assess the

models generalization capability on completely un-

seen data. This set includes difﬁcult, realistic samples

like ﬁsh recorded at difﬁcult visibility conditions and

swimming in and out of the imaged area. While both

models perform comparably well in terms of segmen-

tation accuracy, the PSPNet outperforms DeepLabv3

in regard to inference speed, while achieving an aver-

age pixel accuracy of 96.8% and an intersection-over-

union between the predicted and the target mask of

73.8%.

In the future, we plan to extend this framework for

multi-class segmentation by considering additional

marine species and investigating whether segmenta-

tion results can be improved by considering the en-

tire video sequence instead of individual frames, es-

pecially in cases of difﬁcult visibility. Additionally,

we are going to investigate several reﬁnement steps

that can enhance the segmentation results, such as

Conditional Random Fields (Kr

ahenb

uhl and Koltun,

2011) or afﬁnity matrices (Liu et al., 2017). In a sub-

sequent work, the generated segmentation masks will

be used to extract morphometric features, such as ﬁsh

length, and to locate speciﬁc key points on the ﬁsh

outline. This information will be utilized to deter-

mine the species and biomass of the detected ﬁsh. We

plan to deploy the developed algorithms on an em-

bedded device to be used in remotely operated and au-

tonomous underwater vehicles that complement a net-

work of stationary underwater sensors, with the goal

of performing a continuous ﬁsh stock assessment

REFERENCES

Abbasi, S. and Mokhtarian, F. (1999). Robustness of shape

similarity retrieval under afﬁne transformation. In

Challenge of Image Retrieval, pages 1–10.

Abdeldaim, A. M., Houssein, E. H., and Hassanien, A. E.

(2018). Color image segmentation of ﬁshes with com-

plex background in water. In International Conference

on Advanced Machine Learning Technologies and Ap-

plications, pages 634–643. Springer.

Alsmadi, M. K. and Almarashdeh, I. (2020). A survey on

ﬁsh classiﬁcation techniques. Journal of King Saud

University-Computer and Information Sciences.

Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017).

Segnet: A deep convolutional encoder-decoder ar-

chitecture for image segmentation. IEEE transac-

tions on pattern analysis and machine intelligence,

39(12):2481–2495.

Baloch, A., Ali, M., Gul, F., Basir, S., and Afzal, I. (2017).

Fish image segmentation algorithm (ﬁsa) for improv-

ing the performance of image retrieval system. Inter-

national Journal of Advanced Computer Science and

Applications, 8(12):396–403.

Bicknell, A. W., Godley, B. J., Sheehan, E. V., Votier, S. C.,

and Witt, M. J. (2016). Camera technology for mon-

itoring marine biodiversity and human impact. Fron-

tiers in Ecology and the Environment, 14(8):424–432.

Cavalcanti, M. J., Monteiro, L. R., and Lopes, P. (1999).

Landmark-based morphometric analysis in selected

species of serranid ﬁshes (perciformes: Teleostei).

ZOOLOGICAL STUDIES-TAIPEI-, 38(3):287–294.

Chen, L.-C., Papandreou, G., Schroff, F., and Adam,

H. (2017). Rethinking atrous convolution for

semantic image segmentation. arXiv preprint

arXiv:1706.05587.

Durden, J. M., Schoening, T., Althaus, F., Friedman, A.,

Garcia, R., Glover, A. G., Greinert, J., Stout, N. J.,

Jones, D. O., Jordt, A., et al. (2016). Perspectives in

visual imaging for marine biology and ecology: from

acquisition to understanding. Oceanography and ma-

rine biology: an annual review, 54:1–72.

Fernandes, A. F., Dorea, J. R., and Rosa, G. J. (2020). Im-

age analysis and computer vision applications in an-

Segmentation of Fish in Realistic Underwater Scenes using Lightweight Deep Learning Models

163

imal sciences: an overview. Frontiers in Veterinary

Science, 7:800.

French, G., Fisher, M., Mackiewicz, M., and Needle, C.

(2015). Convolutional neural networks for counting

ﬁsh in ﬁsheries surveillance video.

Garcia, R., Prados, R., Quintana, J., Tempelaar, A., Gracias,

N., Rosen, S., V

agstøl, H., and Løvall, K. (2020). Au-

tomatic segmentation of ﬁsh using deep learning with

application to ﬁsh size measurement. ICES Journal of

Marine Science, 77(4):1354–1366.

Grigorescu, S., Trasnea, B., Cocias, T., and Macesanu,

G. (2020). A survey of deep learning techniques

for autonomous driving. Journal of Field Robotics,

37(3):362–386.

He, K., Gkioxari, G., Doll

ar, P., and Girshick, R. B. (2017).

Mask R-CNN. CoRR, abs/1703.06870.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Konovalov, D. A., Saleh, A., Efremova, D. B., Domingos,

J. A., and Jerry, D. R. (2019). Automatic weight es-

timation of harvested ﬁsh from images. In 2019 Dig-

ital Image Computing: Techniques and Applications

(DICTA), pages 1–7. IEEE.

Kourou, K., Exarchos, T. P., Exarchos, K. P., Karamouzis,

M. V., and Fotiadis, D. I. (2015). Machine learning ap-

plications in cancer prognosis and prediction. Compu-

tational and structural biotechnology journal, 13:8–

17.

ahenb

uhl, P. and Koltun, V. (2011). Efﬁcient inference

in fully connected crfs with gaussian edge potentials.

Advances in neural information processing systems,

24:109–117.

Li, L., Dong, B., Rigall, E., Zhou, T., Donga, J., and Chen,

G. (2021). Marine animal segmentation. IEEE Trans-

actions on Circuits and Systems for Video Technology.

Liu, S., De Mello, S., Gu, J., Zhong, G., Yang, M.-H., and

Kautz, J. (2017). Learning afﬁnity via spatial propa-

gation networks. arXiv preprint arXiv:1710.01020.

Long, J., Shelhamer, E., and Darrell, T. (2015). Fully con-

volutional networks for semantic segmentation. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 3431–3440.

Lopez-Marcano, S., Brown, C. J., Sievers, M., and Con-

nolly, R. M. (2021). The slow rise of technology:

Computer vision techniques in ﬁsh population con-

nectivity. Aquatic Conservation: Marine and Fresh-

water Ecosystems, 31(1):210–217.

Malde, K., Handegard, N. O., Eikvil, L., and Salberg, A.-B.

(2020). Machine intelligence and the data-driven fu-

ture of marine science. ICES Journal of Marine Sci-

ence, 77(4):1274–1285.

Mallet, D. and Pelletier, D. (2014). Underwater video tech-

niques for observing coastal marine biodiversity: a re-

view of sixty years of publications (1952–2012). Fish-

eries Research, 154:44–62.

Marchesan, M., Spoto, M., Verginella, L., and Ferrero,

E. A. (2005). Behavioural effects of artiﬁcial light

on ﬁsh species of commercial interest. Fisheries re-

search, 73(1-2):171–185.

Marmanis, D., Wegner, J. D., Galliani, S., Schindler, K.,

Datcu, M., and Stilla, U. (2016). Semantic segmenta-

tion of aerial images with an ensemble of cnss. ISPRS

Annals of the Photogrammetry, Remote Sensing and

Spatial Information Sciences, 2016, 3:473–480.

Qin, H., Peng, Y., and Li, X. (2014). Foreground extraction

of underwater videos via sparse and low-rank matrix

decomposition. In 2014 ICPR Workshop on Computer

Vision for Analysis of Underwater Imagery, pages 65–

72. IEEE.

Rawat, S., Benakappa, S., Kumar, J., Naik, K., Pandey,

G., and Pema, C. (2017). Identiﬁcation of ﬁsh stocks

based on truss morphometric: A review. Journal of

Fisheries and Life Sciences, 2(1):9–14.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Rosen, S. and Holst, J. C. (2013). Deepvision in-trawl imag-

ing: Sampling the water column in four dimensions.

Fisheries Research, 148:64–73.

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and

Chen, L.-C. (2018). Mobilenetv2: Inverted residu-

als and linear bottlenecks. In Proceedings of the IEEE

conference on computer vision and pattern recogni-

tion, pages 4510–4520.

Spampinato, C., Giordano, D., Di Salvo, R., Chen-Burger,

Y.-H. J., Fisher, R. B., and Nadarajan, G. (2010). Au-

tomatic ﬁsh classiﬁcation for underwater species be-

havior understanding. In Proceedings of the ﬁrst ACM

international workshop on Analysis and retrieval of

tracked events and motion in imagery streams, pages

45–50.

Storbeck, F. and Daan, B. (2001). Fish species recognition

using computer vision and a neural network. Fisheries

Research, 51(1):11–15.

Tchapmi, L., Choy, C., Armeni, I., Gwak, J., and Savarese,

S. (2017). Segcloud: Semantic segmentation of 3d

point clouds. In 2017 international conference on 3D

vision (3DV), pages 537–547. IEEE.

Wilhelms, I. et al. (2013). Atlas of length-weight relation-

ships of 93 ﬁsh and crustacean species from the north

sea and the north-east atlantic. Technical report, Jo-

hann Heinrich von Th

unen Institute, Federal Research

Institute for Rural . . . .

Yang, X., Zeng, Z., Teo, S. G., Wang, L., Chandrasekhar, V.,

and Hoi, S. (2018). Deep learning for practical image

recognition: Case study on kaggle competitions. In

Proceedings of the 24th ACM SIGKDD international

conference on knowledge discovery & data mining,

pages 923–931.

Yu, C., Fan, X., Hu, Z., Xia, X., Zhao, Y., Li, R., and Bai,

Y. (2020). Segmentation and measurement scheme for

ﬁsh morphological features based on mask r-cnn. In-

formation Processing in Agriculture, 7(4):523–534.

Zhao, H., Shi, J., Qi, X., Wang, X., and Jia, J. (2017).

Pyramid scene parsing network. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 2881–2890.

ROBOVIS 2021 - 2nd International Conference on Robotics, Computer Vision and Intelligent Systems

164