TenebrioVision: A Fully Annotated Dataset of Tenebrio Molitor Larvae

Worms in a Controlled Environment for Accurate Small Object

Detection and Segmentation

Angelos-Michael Papadopoulos

, Paschalis Melissas

, Anestis Kastellos

Panagiotis Katranitsiotis

, Panagiotis Zaparas

, Konstantinos Stavridis

and Petros Daras

The Visual Computing Lab - Centre for Research and Technology Hellas/Information Technologies Institute,

Thessaloniki, Greece

Keywords:

TenebrioVision, Dataset, Benchmark, Instance Segmentation, Object Detection, Worms, Tenebrio Molitor,

Edible Insects, Farming, Automation, Alternative Food Source.

Abstract:

Tenebrio molitor worms have shown extreme nutritional beneﬁts, as they contain useful natural compounds,

making them worth as an alternative food source. It is beneﬁcial for insect farms to have automated mech-

anisms that can detect these worms. Without an explicitly annotated dataset, the task of detecting tenebrio

molitor worms remains challenging and underdeveloped. To address this issue, we introduce TenebrioVi-

sion, which is a fully annotated dataset, suitable for the detection and segmentation of tenebrio molitor larvae

worms. The data acquisition is performed in a controlled environment. The dataset consists of 1,120 images,

with a total of 53,600 worm instances. The 1,120 images are equally distributed on 14 distinct levels, each

level containing a speciﬁc number of tenebrio monitor larvae worms. The dataset is validated in terms of

mean average precision, memory allocation, and inference time, on several state-of-the-art baseline methods

for both detection and segmentation purposes. The results unequivocally show that the detection and segmen-

tation accuracy is high on both TenebrioVision and real farm images.

1 INTRODUCTION

With the world’s population expected to reach 9.7 bil-

lion by 2050 (Desa, 2019), the high demand for an-

imal protein, without the detrimental environmental

effects of animal husbandry, poses signiﬁcant chal-

lenges for global food production. The available food

stocks are limited and will eventually become insuf-

ﬁcient to meet this demand. These core factors have

initialized the development of the appropriate indus-

trial production systems (Van Huis et al., 2013).

Insects have a high feed conversion efﬁciency, low

greenhouse gas emissions, high-quality protein, and

require overall fewer resources to produce than other

animal proteins. Moreover, they can be produced on

https://orcid.org/0009-0000-8117-312X

https://orcid.org/0009-0000-2657-9557

https://orcid.org/0009-0002-7812-1577

https://orcid.org/0009-0004-6509-6338

https://orcid.org/0009-0001-5407-7295

https://orcid.org/0000-0002-3244-2511

https://orcid.org/0000-0003-3814-6710

10, 15, 20, ... 80, 90, 100

Figure 1: TenebrioVision: Tenebrio molitor larvae worms

are present in a dataset of 1,120 images, ranging from

frames with 10 worms to frames with 100 worms.

a larger scale than traditional livestock. For these

reasons, several research studies (Ghaly et al., 2009;

Oonincx and De Boer, 2012; Brandon et al., 2021)

Papadopoulos, A., Melissas, P., Kastellos, A., Katranitsiotis, P., Zaparas, P., Stavridis, K. and Daras, P.

TenebrioVision: A Fully Annotated Dataset of Tenebrio Molitor Larvae Worms in a Controlled Environment for Accurate Small Object Detection and Segmentation.

DOI: 10.5220/0012295900003654

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 187-196

ISBN: 978-989-758-684-2; ISSN: 2184-4313

187

have highlighted the beneﬁts of using insects as an

alternative source of animal protein for human con-

sumption.

The yellow mealworm, also known as Tenebrio

molitor, has demonstrated signiﬁcant potential for hu-

man consumption. There are four development stages

of this species: Eggs, larvae, pupae, and adults. Tene-

brio molitor larvae is typically the preferred stage in

many countries where insects are consumed (Stoops

et al., 2016; Siemianowska et al., 2013). It is a valu-

able source of protein and minerals, easy to main-

tain, and can be harvested at an earlier stage of de-

velopment. Several studies (Costa et al., 2020) and

(Kr

oncke and Benning, 2022) have shown that Tene-

brio molitor includes a nutritional composition sim-

ilar to other conventional meat sources. Therefore,

the European Commission implementing regulation

2023/58 (of the European Union, 2023) has ofﬁcially

authorized the placing on the market of frozen, paste,

dried, and powdered larvae of Alphitobius diaperinus

(minor mealworm) as a novel food.

Tenebrio molitor, in its larvae development stage,

is therefore becoming increasingly popular for farm-

ing and raising edible insects, which highlights the

need for standardized and cost-effective production

techniques. There are already insect farms oper-

ating in numerous countries, but their production

systems lack efﬁcient and scalable automation pro-

cesses (Grau et al., 2017), which can be supported by

computer vision and machine learning technologies.

Tenebrio molitor breeding procedures, such as feed-

ing, wetting larvae, classifying larvae by size, harvest-

ing chitinous moult, and ﬁnally harvesting larvae and

separating them from impurities, should be automated

and monitored in order to be proﬁtable. The amount

of manual labor required by current farming methods

prevents them from being used for industrial-scale

production.

In this paper, a publicly available dataset called

“TenebrioVision” is introduced, which is a compre-

hensive, and fully annotated dataset for Tenebrio

Molitor larvae insects in a controlled environment.

TenebrioVision dataset consists of 1,120 images of

53,600 tenebrio molitor instances annotated for both

object detection and instance segmentation purposes

(ﬁgure 1). Every image is taken in a controlled setup

by a UI camera (UI-3884LE, 2021). The resolution of

each frame is 3088 x 2076 pixels. The dataset is dis-

tributed in 14 different levels, depending on the num-

ber of tenebrio molitor instances inside the crate, as

shown in ﬁgure 4. It should be noticed that localizing

insects in the visual scene at the level of the instance

mask is a fundamental step that enables further qual-

ity analysis in the ﬁeld. The signiﬁcance of Tenebri-

oVision dataset is then validated by running several

state-of-the-art baseline methods for detection and in-

stance segmentation. The intention for experimenting

with several SoTA models is to compare both the de-

tection/segmentation efﬁciency, but also their infer-

ence time on the TenebrioVision dastaset. A quick

inference time can be a vital asset in the farm indus-

tries, thus a comprehensive comparison is necessary.

Lastly, the high-value features learned from Tenebri-

oVision are tested on farm images, taken from a vari-

ety of real farm industry environments, some of them

containing a huge amount of tenebrio molitor insects,

uncountable even by an insect expert’s human eye.

By making this dataset available to the public, it is

believed that the following aspects will be positively

impacted:

• To the best of our knowledge, the TenebrioVision

dataset is the ﬁrst fully annotated dataset that in-

cludes the tenebrio molitor larvae insects for de-

tection/segmentation tasks at a large scale.

• The TenebrioVision dataset can address chal-

lenges associated with automated insect breeding

and production (Cadinu et al., 2020).

• It can also serve as a small object detection/seg-

mentation benchmark for researchers and indus-

try professionals for a variety of computer vision

tasks.

• Rich features learned by state-of-the-art mod-

els from the augmented TenebrioVision dataset,

proved to be substantial in real-case scenarios,

surpassing expert human-eye capabilities.

2 RELATED WORK

2.1 Insect Related Datasets

Popular datasets like ImageNet (Deng et al., 2009)

and COCO (Lin et al., 2014) that include different

classes of animals have been developed over the years

for a variety of computer vision tasks, such as im-

age classiﬁcation and object detection - segmenta-

tion, respectively. The interest in building specialized

datasets that are only concerned with animal species

has grown in recent years. Datasets like iNaturalist

(Van Horn et al., 2018), which currently contains over

415,000 species of animals, Animal Kingdom dataset

(Ng et al., 2022) containing 850 species across 6 ma-

jor animal classes, and others (Beery et al., 2021;

Gagne et al., 2021; Cao et al., 2019) have a major con-

tribution to analyzing animal behavior. Some datasets

are devoted exclusively to one or a few animal species

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

188

Table 1: Comparison of various domain-speciﬁc datasets.

Task

Dataset Worm Images

Tenebrio molitor larvae

Images Classiﬁcation Detection Segmentation

iNaturalist(Van Horn et al., 2018) 436,200 1,620 X - -

IP102(Wu et al., 2020) 6,850 - X X -

Larvae Dataset in roboﬂow(Probst, 2023) 179 179 -

No. instances: 179

Mealworms Dataset in roboﬂow(egg detection mixed eggs, 2022) 518 518 X

No. instances: 518

Multipurpose monitoring system(Majewski et al., 2022) 120 120 X X

No. instances: 1,026

TenebrioVision 1,120 1,120 - X

No. instances: 53,600

only (Wah et al., 2011; Fang et al., 2020; Labuguen

et al., 2021; Nuthalapati and Tunga, 2021)

Regarding solely insects, there are datasets like

(Van Horn et al., 2018; Hansen et al., 2020; Wu et al.,

2020) that include large amounts of insect images.

Yet, there have not been many attempts to study in-

sect datasets resembling tenebrio molitor worms in

the larvae stage, which is the nutrition-preferred one

as the ingredient for the EU foods. Efforts have been

made by (Hebert et al., 2021) to create synthetic im-

ages of worm posture, which would avoid the need for

human-labeled annotation. Other researchers (Hus-

son et al., 2018) and (Pereira et al., 2019) have de-

veloped tools for analyzing their pose and behavior.

However, due to the nature and characteristics of real

tenebrio-molitor worms, these approaches cannot be

effective in real-world schemes.

Furthermore, regardless of whether Tenebrio

molitor is present at its larvae stage or not, all of

the existing datasets contain a small amount of the

particular tenebrio molitor and are primarily used

for animal classiﬁcation tasks. The development of

techniques for examining Tenebrio Molitor’s general

characteristics is severely hampered by this barrier.

Thus, the need for tenebrio molitor images for both

image detection and segmentation tasks emerges.

There are two small, yet preferable datasets regard-

ing tenebrio molitor worms in the larvae development

stage (egg detection mixed eggs, 2022) and (Probst,

2023). The ﬁrst one classiﬁes the images of live or

dead tenebrio molitor and also detects the worms.

Both of them can be found in Roboﬂow (Dwyer, ).

However, they feature only a limited sample of tene-

brio molitor images. These datasets are annotated

solely for object detection and typically showcase just

a single insect per frame, failing to represent authen-

tic farming conditions. In a recent paper (Majewski

et al., 2022), the authors made a multipurpose mon-

itoring system for Tenebrio molitor breeding. Tene-

brio molitor larvae, pupae, and beetles are all de-

tected by the instance segmentation module (ISM),

which also detects dead larvae and pests as anoma-

lies. From the acquired image, the semantic segmen-

tation module (SSM) extracted feed, chitin, and frass.

Additionally, the Larvae Phenotyping Module (LPM)

computes features for both the population as a whole

and each individual larva (length, curvature, mass,

segmentation, and classiﬁcation). However, due to

the difﬁcult process of annotation and the need for

a multipurpose monitoring system, they present only

120 total labeled images, including 1026 live tenebrio

molitor larvae instances among others. These images

do not all contain the tenebrio molitor insect in the

larval stage. Also, these images are not publicly avail-

able.

A comparison of the aforementioned domain-

speciﬁc datasets is presented in table 1. For each

dataset, the total number of worm images, the total

number of tenebrio molitor worm images and the pur-

pose that this dataset accomplishes(classiﬁcation, de-

tection, segmentation) are provided. It is shown that

even though there are large datasets that contain many

worm images (iNaturalist (Van Horn et al., 2018),

IP102 (Wu et al., 2020)) there is a limited number

of tenebrio molitor images on the larvae development

stage. On top of that, all these datasets either contain

limited (Majewski et al., 2022) or no data at all for the

segmentation task.

Given the aforementioned factors, it becomes ev-

ident that a direct and equitable comparison between

these datasets and TenebrioVision, in terms of perfor-

mance, is not feasible.

TenebrioVision: A Fully Annotated Dataset of Tenebrio Molitor Larvae Worms in a Controlled Environment for Accurate Small Object

Detection and Segmentation

189

2.2 Vision Based Methods for Insect

Analysis

Large strides have been made in the ﬁeld of object de-

tection in images, particularly with algorithms based

on deep learning, which are typically divided into

two categories: Two-stage detectors (methods based

on the region proposal network), like Faster R-CNN

(Girshick, 2015) and Mask R-CNN (He et al., 2017),

and one-stage detectors like YOLO (Redmon et al.,

2016), SSD (Liu et al., 2016).

Various models such as Mask R-CNN (He et al.,

2017) and YOLO-V5 (Jocher, 2020) have been ex-

plored by researchers for the detection and segmen-

tation of worms (Majewski et al., 2022). However,

given the speciﬁc requirements of accuracy and rapid

inference time in the context of insect farms, an exten-

sive evaluation of ﬁve state-of-the-art baseline mod-

els is conducted in the TenebrioVision dataset. These

models included Mask R-CNN (He et al., 2017), Ef-

ﬁcientDet (Tan et al., 2020), YOLO-V7 (Wang et al.,

2022), YOLO-V8 (Jocher et al., tics) and the current

state-of-the-art YOLO-NAS (Aharon et al., 2021).

3 TenebrioVision DATASET

Detailed information about the TenebrioVision

dataset, which comprises a large collection of high-

quality images of Tenebrio molitor larvae worms in

various poses and orientations are discussed in detail

in this section.

3.1 Experimental Setup

The TenebrioVision dataset is collected using a

custom-designed setup consisting of a crate in a con-

trolled environment and a UI-3884LE-C-HQ camera

(UI-3884LE, 2021) placed above it. The UI camera is

chosen because it offers 3088 x 2076 pixel resolution

at frame rates up to 58.0 fps even under low-light con-

ditions and the focus can be conveniently adjusted.

The tenebrio molitor worms are placed inside a crate

with a spatial ﬁeld of view of 20cm x 30cm, which

is exactly a quarter of the crate’s spatial ﬁeld of view,

as it is presented in ﬁgure 2. The quarter of the crate,

as a spatial ﬁeld of view, of the UI camera was cho-

sen experimentally in order to capture the 3088x2076

pixel resolution. By taking into account the following

equation (1) from (Fulton, 2015), it is concluded that

38 cm is the ideal height to set the camera from the

crate:

Figure 2: Experimental setup. The UI camera is placed at a

38 cm distance from the black crate, according to Equation

1. The spatial ﬁeld of view is 20 x 30 cm. The camera is

steadily placed on the desk during the whole experimental

procedure.

D(cm) =

Q × F

(1)

where D is the camera’s distance from the crate

in cm, Q is the Quarter crate’s size: 20cm x 30cm,

F is the Focal length of the camera: 9.6 mm and the

H is the Object height on the sensor, which is actu-

ally the optical size of the camera sensor: 7.411mm

x 4.982mm according to UI-3884LE-C-HQ speciﬁca-

tions (UI-3884LE, 2021).

3.2 Dataset Acquisition

The Tenebrio molitor larvae worms are placed inside

the crate and allowed to move freely for a few seconds

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

190

(a) (b)

Figure 3: TenebrioVision dataset statistics. (a) Train, validation and test split of TenebrioVision dataset with respect to both

the total number of instances and the total number of images. (b) This pie - chart portrays the distribution of TenebrioVision’s

levels among the entire dataset.

while being captured by the UI camera. More partic-

ularly, a 30-second video is recorded and processed

to only obtain the ﬁrst and last frames. This choice is

made because worm position and orientation between

the ﬁrst and last images are likely to differ signiﬁ-

cantly ensuring that there are no identical images in

the dataset, thus providing no bias. When the desired

number of images has been obtained, the process is

complete.

In order to capture the ground truth illumination

inside the crate, which is crucial for accurately repre-

senting the colors of the worms, the speciﬁcations of

the UI camera are modiﬁed. Speciﬁcally, the sensi-

tivity is increased and the white balance settings are

adjusted for the purpose of capturing the true colors

of the larvae.

In order to guarantee that only live Tenebrio moli-

tor larvae are included in the dataset throughout the

experimental process, the worm specimens are con-

sistently provided with adequate nutrition and temper-

ature control. This procedure enables the collection of

rich, representative data so that the deep learning ap-

proaches will learn features associated only with live

tenebrio molitor worms.

The captioning framework is initialized after the

camera setup. In order to have a balanced distribution

of the images across the dataset, 14 levels are formed,

according to the number of worms inside each image.

Each level contains 80 images. There are 14 levels:

10, 15, 20, 25, 30, 35, 40, 45, 50, 60, 70, 80, 90, and

100 as it is demonstrated in the ﬁgure 4. In essence,

level 45 has 80 images, and each image includes 45

worm instances. Figure 3b depicts the proportion of

each level to the total number of worm instances in the

TenebrioVision dataset. In total, the TenebrioVision

dataset includes 1,120 images, and each image has a

resolution of 6 Megapixels.

3.3 Dataset Annotation

Annotation is performed using Datatorch (Nguyen,

2020) by drawing precise segmentation masks cov-

ering each tenebrio molitor worm in all frames. Tene-

brio molitor larvae’s high level of articulation and po-

tential occlusion in crowded environments lead to an

increased annotation effort. Without accounting for

the overhead, annotating 1,120 images with segmen-

tations masks and bounding boxes for the TenebrioVi-

sion dataset took more than 373 human hours of an-

notation time. The dataset is publicly available here:

https://vcl.iti.gr/dataset/TenebrioVision/.

3.4 Dataset Split

TenebrioVision. To enable the reproduction of the

experimental results, a 75:15:10 split is followed, for

the training, validation, and test set, respectively, as

shown in ﬁgure 3a.

TenebrioVision: A Fully Annotated Dataset of Tenebrio Molitor Larvae Worms in a Controlled Environment for Accurate Small Object

Detection and Segmentation

191

(a) level 10 (b) level 15 (c) level 20 (d) level 25

(e) level 30 (f) level 35 (g) level 40 (h) level 45

(i) level 50 (j) level 60 (k) level 70 (l) level 80

(m) level 90 (n) level 100

Figure 4: Levels of the TenebrioVision dataset. There are 14 levels. The numerical value assigned to each level indicates the

cumulative count of tenebrio molitor worm instances captured in the corresponding images at that level. Each level has 80

images, totaling 1,120 images. For example, we have 80 images for level 90, and every image at level 90 contains only 90

tenebrio monitor worms.

4 EXPERIMENT AND

EVALUATION

In this section, the performance of TenebrioVision

dataset is evaluated on various state-of-the-art object

detection and instance segmentation models. The im-

ages in the training set are resized to each baseline

setup, always trying, if possible, to maintain the as-

pect ratio of 6 Megapixels. The performance of each

model is evaluated based on mean average precision

(mAP), inference time, and memory allocation. These

evaluation metrics are chosen because they are essen-

tial indicators of the ﬁnal product’s quality in auto-

mated farming. Finally, a qualitative evaluation of

images provided by real insect farms is presented. All

experiments are conducted on the NVIDIA GeForce

RTX 3090 GPU with 24GB memory.

4.1 Quantitative Evaluation

The ﬁve baselines baseline models used are Mask R-

CNN (He et al., 2017), EfﬁcientDet (Tan et al., 2020),

YOLO-V7 (Wang et al., 2022) and YOLO-V8 (Jocher

et al., tics) for both object detection (OD) and instance

segmentation (IS) tasks and the YOLO-NAS (Aharon

et al., 2021). Mask R-CNN is conducted using the

Detectron2 (Wu et al., 2019) framework. Detectron2

is preferred over the original implementation (Mat-

terport Mask R-CNN (Abdulla, 2017)) due to its im-

proved performance and ﬂexibility. The implementa-

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

192

(a) object detection for a

level 35 image

(b) object detection for a

level 50 image

level 70 image

(d) object detection for a

level 100 image

(e) instance segmentation

for a level 40 image

(f) instance segmentation for

a level 45 image

(g) instance segmentation

for a level 70 image

(h) instance segmentation

for a level 80 image

Figure 5: Experiment results on TenebrioVision’s dataset test set. On top, there are predictions regarding object detection

on different levels of the dataset, and below there are predictions regarding instance segmentation on different levels of the

dataset.

Table 2: Detection and segmentation results on the test set

of TenebrioVision using different state-of-the-art models.

We evaluate the mAP@0.5:0.05:0.95 and mAP@0.75 for

the bounding box and the mask. Empty objects on the table

indicates that the model performs only object detection.

Methods mAP bbox

mAP@75 bbox

mAP mask

mAP@75 mask

Mask RCNN 0.781 0.954 0.654 0.776

EFFICIENTDET 0.787 0.926 - -

YOLO V7 OD 0.842 0.93 - -

YOLO V7 IS 0.848 0.93 0.632 0.802

YOLO V8 OD 0.874 0.96 - -

YOLO V8 IS 0.88 0.965 0.729 0.83

YOLO NAS 0.892 0.972 - -

tion of the other baselines is taken from their ofﬁcial

repositories.

For each method, the smallest backbone version

of the model is used. For Mask R-CNN the backbone

is Resnet-50-FPN, for EfﬁcientDet the EfﬁcientDet-

D0, for Yolov7-(OD) the yolov7-tiny, for Yolov7-

(IS) yolov7-seg, for yolov8-(OD) the yolov8n, for

Yolov8-(IS) yolov8n-seg and ﬁnally for Yolo-NAS

the YOLO-NAS-s. The reason that the largest back-

bones of those baselines are not employed is twofold:

ﬁrstly, there is no discernible difference in the output

mAP, and secondly, inference time decreases. A basic

augmentation scheme is followed for these baselines.

The basic augmentations are Horizontal and Vertical

ﬂips. For Mask R-CNN and Yolo-V7 the training im-

Table 3: Comparison of memory allocation in Giga-

bytes(GBs) and inference time (FPS) during the testing

phase on several state-of-the-art models.

Methods Memory(GBs) FPS

Faster R-CNN 2.27 20

Mask R-CNN 5.06 9

EfﬁcientDet 1.75 25

YOLO V7 OD 2.37 220

YOLO V7 IS 3.44 40

YOLO V8 OD 1.59 150

YOLO V8 IS 2.12 80

YOLO NAS 1.48 416

age has a 1024 x 688 resolution with respect to the

camera’s aspect ratio. For EfﬁcientDet the training

image size is 512 x 512 in order to get the small-

est EfﬁcientDet version (EfﬁcientDet-D0). Finally,

the training image size for Yolo-V8 and Yolo-NAS

is 1280 x 1280.

The results are presented in table 2, where the

mAP scores for all methods tested are presented. In

table 3 the memory allocation and the inference time

of each method are reported. The memory allocation

refers to the Gigabytes (GBs) required to infer an im-

age from the TenebrioVision dataset and the inference

time to the frames per second (FPS). The optimal per-

TenebrioVision: A Fully Annotated Dataset of Tenebrio Molitor Larvae Worms in a Controlled Environment for Accurate Small Object

Detection and Segmentation

193

Figure 6: Extended augmentation scheme for TenebrioVi-

sion dataset. The applied augmentations are horizontal-

vertical ﬂips, random rotation, adding noise, zooming-in,

brightness, and exposure. The annotated segmentation

masks are also presented.

formance is achieved using the current SoTA Yolo-

V8(IS) (Jocher et al., tics) with a mAP of 0.729 for the

segmentation mask and the Yolo-NAS (Aharon et al.,

2021) with a mAP of 0.892 for the bounding box, both

evaluated at IoU=0.5:.05:0.95. Considering the infer-

ence time, the YOLO-NAS model exhibits the low-

est memory allocation of 1.48 GB for object detec-

tion, while YOLO-V8 (IS) requires 2.12 GB for the

instance segmentation task. Additionally, the highest

FPS values of 416 and 80 for detection and segmenta-

tion tasks are attained by Yolo-NAS and Yolo-V8(IS)

respectively, on the TenebrioVision dataset.

4.2 Qualitative Evaluation

Some experimental results regarding object detection

and instance segmentation can be seen in ﬁgure 5.

This ﬁgure clearly shows that small and highly articu-

lated objects like those of tenebrio molitor can be ac-

curately detected. A precise segmentation mask can

provide a wealth of information about these worms.

Information like color, length, width and size of the

tenebrio molitor larvae can help experts gain knowl-

edge about the health and life cycle of each worm.

This analysis is really vital for farm workers as it

saves them a lot of time when trying to ﬁgure out if

a crate ﬁlled with worms is healthy or unhealthy in a

production pipeline scheme. These data will aid the

competent scientiﬁc community and hasten the breed-

ing of insects for insect farming.

Further evaluation of the models is conducted on

actual farm images, which contain only tenebrio moli-

tor worms at the larvae stage. It is found that even

(a) Ground Truth (b) Basic Aug-

mentations

Augmentations

(d) Ground Truth (e) Basic Aug-

mentations

(f) Extended

Augmentations

(g) Ground Truth (h) Basic Aug-

mentations

(i) Extended

Augmentations

(j) Ground Truth (k) Basic Aug-

mentations

(l) Extended

Augmentations

Figure 7: Expreriments conducted on real farm images. The

left column depicts the Ground Truth farm images, the cen-

ter column the predictions of the SoTA models, trained on

TenebrioVision with Basic Augmentations, and the right

column the prediction of the SoTA models, trained on Tene-

brioVision with an extended augmentation scheme.

though the current SoTA baseline methods achieved

a satisfying mAP at the test set, their performance

on real farm images is poor and inadequate as it is

demonstrated in ﬁgure 7 (center column). Genuine

farm images, in contrast to TenebrioVision, show a

signiﬁcant amount of Tenebrio molitor larvae inside

the inspection crates, frequently beyond the limit of

human counting capacity. Real farm images typically

have a range of backgrounds, colors, camera angles,

and many other aspects. To tackle this problem the

best models are trained with an extended augmenta-

tion scheme. More speciﬁcally, these augmentations

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

194

are horizontal-vertical ﬂips, random rotation, adding

noise, zooming-in and out, brightness, and exposure.

The noise augmentation technique applies a Gaus-

sian distribution to 8% of the pixels in each image.

Samples from this extended augmentation scheme are

demonstrated in ﬁgure 6. The results from this ex-

tended augmentation scheme are fascinating, as seen

in ﬁgure 7 (right column), since YOLO-V8(IS) now

detects and captures a very large amount of tenebrio

molitor larvae worms, that are even uncountable by

an expert’s human eye.

It is believed that the process of inference on real

images and adding them back to the TenebrioVision

will further improve the detection/segmentation tasks,

and enhance the automation process of the farms.

5 CONCLUSIONS

In this paper, TenebrioVision was introduced, a

dataset that contains tenebrio molitor worms in

the larvae development stage. The TenebrioVision

dataset contains 1,120 fully annotated images for de-

tection and segmentation tasks, at a resolution of

3088x2076 pixels. The total number of worm in-

stances is 53,600. The performance of several state-

of-the-art object detection and instance segmentation

models was evaluated on TenebrioVision. The exper-

iments’ ﬁndings demonstrate that, despite the SoTA

algorithms’ robustness, they perform poorly in real-

world situations, necessitating augmentations. By ex-

tending TenebrioVision with augmentations, aston-

ishing results were achieved, surpassing the expert

human-eye detection accuracy. By making Tenebri-

oVision publicly available our aim is to assist and en-

hance the automation process on real farms, thus help-

ing the scientiﬁc community produce valuable data

and knowledge on this high-nutrition worm. This

dataset can also be utilized, by the computer vision

community, as a benchmark for small object detec-

tion and segmentation tasks.

ACKNOWLEDGMENTS

The work presented in this paper was supported by

the European Commission under contract H2020 -

101016953 CoRoSect.

REFERENCES

Abdulla, W. (2017). Mask r-cnn for object detection and

instance segmentation on keras and tensorﬂow. https:

//github.com/matterport/Mask RCNN.

Aharon, S., Louis-Dupont, Ofri Masad, Yurkova, K., Lotem

Fridman, Lkdci, Khvedchenya, E., Rubin, R., Bagrov,

N., Tymchenko, B., Keren, T., Zhilko, A., and Eran-

Deci (2021). Super-gradients.

Beery, S., Agarwal, A., Cole, E., and Birodkar, V. (2021).

The iwildcam 2021 competition dataset. arXiv

preprint arXiv:2105.03494.

Brandon, A. M., Garcia, A. M., Khlystov, N. A., Wu, W.-

M., and Criddle, C. S. (2021). Enhanced bioavailabil-

ity and microbial biodegradation of polystyrene in an

enrichment derived from the gut microbiome of tene-

brio molitor (mealworm larvae). Environmental sci-

ence & technology, 55(3):2027–2036.

Cadinu, L. A., Barra, P., Torre, F., Delogu, F., and Madau,

F. A. (2020). Insect rearing: Potential, challenges, and

circularity. Sustainability, 12(11):4567.

Cao, J., Tang, H., Fang, H.-S., Shen, X., Lu, C., and Tai, Y.-

W. (2019). Cross-domain adaptation for animal pose

estimation. In Proceedings of the IEEE/CVF Interna-

tional Conference on Computer Vision, pages 9498–

9507.

Costa, S., Pedro, S., Lourenc¸o, H., Batista, I., Teixeira, B.,

Bandarra, N. M., Murta, D., Nunes, R., and Pires, C.

(2020). Evaluation of tenebrio molitor larvae as an

alternative food source. NFS journal, 21:57–64.

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-

Fei, L. (2009). Imagenet: A large-scale hierarchical

image database. In 2009 IEEE Conference on Com-

puter Vision and Pattern Recognition, pages 248–255.

Desa, U. (2019). World population prospects 2019: High-

lights. New York (US): United Nations Department for

Economic and Social Affairs, 11(1):125.

Dwyer, B., N. J. . S. J. e. a. R. V. . S. A. f. u.

egg detection mixed eggs (2022). Mealworms-

detection dataset. https://universe.roboﬂow.com/

egg-detection-mixed-eggs/mealworms-detection.

visited on 2023-02-23.

Fang, C., Zhang, T., Zheng, H., Huang, J., and Cuan, K.

(2020). Pose estimation and behavior classiﬁcation of

broiler chickens based on deep neural networks. Com-

put. Electron. Agric., 180:105863.

Fulton, W. (2015). Calculate distance or size of an object

in a Photo Image. https://www.scantips.com/lights/

subjectdistance.html.

Gagne, C., Kini, J., Smith, D., and Shah, M. (2021).

Florida wildlife camera trap dataset. arXiv preprint

arXiv:2106.12628.

Ghaly, A. E., Alkoaik, F., et al. (2009). The yellow meal-

worm as a novel source of protein. American Jour-

nal of Agricultural and Biological Sciences, 4(4):319–

331.

Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE

international conference on computer vision, pages

1440–1448.

Grau, T., Vilcinskas, A., and Joop, G. (2017). Sustain-

able farming of the mealworm tenebrio molitor for the

production of food and feed. Zeitschrift f

ur Natur-

forschung C, 72(9-10):337–349.

Hansen, O. L., Svenning, J.-C., Olsen, K., Dupont, S.,

Garner, B. H., Iosiﬁdis, A., Price, B. W., and Høye,

TenebrioVision: A Fully Annotated Dataset of Tenebrio Molitor Larvae Worms in a Controlled Environment for Accurate Small Object

Detection and Segmentation

195

T. T. (2020). Species-level image classiﬁcation with

convolutional neural network enables insect identiﬁ-

cation from habitus images. Ecology and Evolution,

10(2):737–747.

He, K., Gkioxari, G., Doll

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

Mask r-cnn. In Proceedings of the IEEE international

conference on computer vision, pages 2961–2969.

Hebert, L., Ahamed, T., Costa, A. C., O’Shaughnessy,

L., and Stephens, G. J. (2021). Wormpose: Image

synthesis and convolutional networks for pose esti-

mation in c. elegans. PLoS computational biology,

17(4):e1008914.

Husson, S. J., Costa, W. S., Schmitt, C., and Gottschalk, A.

(2018). Keeping track of worm trackers. WormBook:

The Online Review of C. elegans Biology [Internet].

Jocher, G. (2020). YOLOv5 by Ultralytics.

Jocher, G., Chaurasia, A., and Qiu, J. (2023., {https://

github.com/ultralytics/ultralytics}). Yolo by ultralyt-

ics.

oncke, N. and Benning, R. (2022). Self-selection of feed-

ing substrates by tenebrio molitor larvae of different

ages to determine optimal macronutrient intake and

the inﬂuence on larval growth and protein content. In-

sects, 13(7):657.

Labuguen, R., Matsumoto, J., Negrete, S. B., Nishimaru,

H., Nishijo, H., Takada, M., Go, Y., Inoue, K.-i., and

Shibata, T. (2021). Macaquepose: a novel “in the

wild” macaque monkey pose dataset for markerless

motion capture. Frontiers in behavioral neuroscience,

14:581154.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., Doll

ar, P., and Zitnick, C. L. (2014).

Microsoft coco: Common objects in context. In Com-

puter Vision–ECCV 2014: 13th European Confer-

ence, Zurich, Switzerland, September 6-12, 2014, Pro-

ceedings, Part V 13, pages 740–755. Springer.

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.,

Fu, C.-Y., and Berg, A. C. (2016). Ssd: Single shot

multibox detector. In Computer Vision–ECCV 2016:

14th European Conference, Amsterdam, The Nether-

lands, October 11–14, 2016, Proceedings, Part I 14,

pages 21–37. Springer.

Majewski, P., Zapotoczny, P., Lampa, P., Burduk, R., and

Reiner, J. (2022). Multipurpose monitoring system

for edible insect breeding based on machine learning.

Scientiﬁc Reports, 12(1):1–15.

Ng, X. L., Ong, K. E., Zheng, Q., Ni, Y., Yeo, S. Y., and

Liu, J. (2022). Animal kingdom: A large and di-

verse dataset for animal behavior understanding. In

Proceedings of the IEEE/CVF Conference on Com-

puter Vision and Pattern Recognition (CVPR), pages

19023–19034.

Nguyen, M. (2020). Datatorch annotation tool. https:

//datatorch.io/.

Nuthalapati, S. V. and Tunga, A. (2021). Multi-domain few-

shot learning and dataset for agricultural applications.

In Proceedings of the IEEE/CVF International Con-

ference on Computer Vision, pages 1399–1408.

of the European Union, O. J. (2023). Commission im-

plementing regulation (eu) 2023/58. https://eur-lex.

europa.eu/eli/reg impl/2023/58/oj. Ofﬁcial Journal of

the European Union, 66.

Oonincx, D. G. and De Boer, I. J. (2012). Environmental

impact of the production of mealworms as a protein

source for humans–a life cycle assessment. PloS one,

7(12):e51145.

Pereira, T. D., Aldarondo, D. E., Willmore, L., Kislin,

M., Wang, S. S.-H., Murthy, M., and Shaevitz, J. W.

(2019). Fast animal pose estimation using deep neural

networks. Nature methods, 16(1):117–125.

Probst, A. (2023). larvae dataset. https://universe.

roboﬂow.com/andre-probst-egt9g/larvae-bbsvy. vis-

ited on 2023-02-23.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time object

detection. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 779–

788.

Siemianowska, E., Kosewska, A., Aljewicz, M., Skib-

niewska, K. A., Polak-Juszczak, L., Jarocki, A., and

Jedras, M. (2013). Larvae of mealworm tenebrio moli-

tor l. as european novel food.

Stoops, J., Crauwels, S., Waud, M., Claes, J., Lievens, B.,

and Van Campenhout, L. (2016). Microbial commu-

nity assessment of mealworm larvae (tenebrio moli-

tor) and grasshoppers (locusta migratoria migratori-

oides) sold for human consumption. Food Microbiol-

ogy, 53:122–127.

Tan, M., Pang, R., and Le, Q. V. (2020). Efﬁcientdet: Scal-

able and efﬁcient object detection. In Proceedings

of the IEEE/CVF conference on computer vision and

pattern recognition, pages 10781–10790.

UI-3884LE (2021). Ids. ui-3884le-c-hq camera (ab00979).

https://en.ids-imaging.com/IDS/datasheet pdf.php?

sku=AB00979.

Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun,

C., Shepard, A., Adam, H., Perona, P., and Belongie,

S. (2018). The inaturalist species classiﬁcation and

detection dataset. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 8769–8778.

Van Huis, A., Van Itterbeeck, J., Klunder, H., Mertens, E.,

Halloran, A., Muir, G., and Vantomme, P. (2013). Ed-

ible insects: future prospects for food and feed secu-

rity. Number 171. Food and agriculture organization

of the United Nations.

Wah, C., Branson, S., Welinder, P., Perona, P., and Be-

longie, S. (2011). The caltech-ucsd birds-200-2011

dataset.

Wang, C.-Y., Bochkovskiy, A., and Liao, H.-Y. M. (2022).

Yolov7: Trainable bag-of-freebies sets new state-of-

the-art for real-time object detectors. arXiv preprint

arXiv:2207.02696.

Wu, X., Zhan, C., Lai, Y.-K., Cheng, M.-M., and Yang,

J. (2020). Ip102: A large-scale benchmark dataset

for insect pest recognition. In Proceedings of the

IEEE/CVF conference on computer vision and pattern

recognition, pages 8787–8796.

Wu, Y., Kirillov, A., Massa, F., Lo, W.-Y., and Gir-

shick, R. (2019). Detectron2. https://github.com/

facebookresearch/detectron2.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

196