Animal Fiber Identiﬁcation under the Open Set Condition

Oliver Rippel

1 a

, Sergen G

ulc¸elik

, Khosrow Rahimi

, Juliana Kurniadi

, Andreas Herrmann

and Dorit Merhof

1 b

Institute of Imaging & Computer Vision, RWTH Aachen University, Aachen, Germany

DWI – Leibniz-Institut f

ur Interaktive Materialien, Aachen, Germany

Keywords:

Out-of-Distribution Detection, Natural Fiber Identiﬁcation, Classiﬁcation, Open Set Recognition, Machine

Learning.

Abstract:

Animal ﬁber identiﬁcation is an essential aspect of fabric production, since specialty ﬁbers such as cashmere

are often targeted by adulteration attempts. Proposed, automated solutions can furthermore not be applied in

practice (i.e. under the open set condition), as they are trained on a small subset of all existing ﬁber types only

and simultaneously lack the ability to reject ﬁber types unseen during training at test time. In our work, we

overcome this limitation by applying out-of-distribution (OOD)-detection techniques to the natural ﬁber iden-

tiﬁcation task. Speciﬁcally, we propose to jointly model the probability density function of in-distribution data

across feature levels of the trained classiﬁcation network by means of Gaussian mixture models. Moreover,

we extend the open set F-measure to the so-called area under the open set precision-recall curve (AUPR

), a

threshold-independent measure of joint in-distribution classiﬁcation & OOD-detection performance for OOD-

detection methods with continuous OOD scores. Exhaustive comparison to the state of the art reveals that our

proposed approach performs best overall, achieving highest area under the class-averaged, open set precision-

recall curve (AUPR

os,avg

). We thus show that the application of automated ﬁber identiﬁcation solutions under

the open set condition is feasible via OOD detection.

1 INTRODUCTION

Animal ﬁbers possess desirable characteristics such

as thermal insulation, moisture wicking and softness,

making them an important material for fabric pro-

duction (McGregor, 2018). Since specialty ﬁbers

(e.g. cashmere) excel at one or more of the above

properties they achieve premium prices on the mar-

ket (International Wool Textile Organisation, 2018).

Said prices, however, render specialty ﬁbers an at-

tractive target for adulteration, and adulteration rates

between 15-60% have been reported for cashmere

products (Waldron et al., 2014; Phan and Wortmann,

2001).

Various ﬁber identiﬁcation methods have been

proposed to counteract adulteration (Rane and Barve,

2011; Kim et al., 2013; Zoccola et al., 2013; In-

ternational Wool Textile Organisation, 2000; Amer-

ican Society for Testing and Materials, 1993). Out

of those, optical identiﬁcation methods are the

most widely applicable. Here, ﬁbers are identi-

https://orcid.org/0000-0002-4556-5094

https://orcid.org/0000-0002-1672-2185

ﬁed based on their surface morphology using ei-

ther optical microscopy (American Society for Test-

ing and Materials, 1993) or scanning electron mi-

croscopy (SEM) (International Wool Textile Organ-

isation, 2000). Despite being subjective in nature and

requiring extensively trained experts to achieve reli-

able results (Zhang and Ainsworth, 2005; Wortmann,

1991), optical identiﬁcation methods are still the ones

predominantly used in industry.

In order to overcome the limitations of human ex-

perts, it has been proposed to automate the ﬁber iden-

tiﬁcation by means of pattern recognition (Yildiz,

2020; Robson, 1997; Robson, 2000; Xing et al.,

2020a; Xing et al., 2020b; Rippel et al., 2021a).

While prior work has shown that accurate ﬁber iden-

tiﬁcation is possible, the application of the developed

solutions in practice is hindered by the following two

facts: (I) Developed solutions train & evaluate their

algorithms only on a small subset of all existing ﬁber

types. Even though up to 11 ﬁber types (10 specialty

ﬁbers + wool) can be identiﬁed by a human expert

(International Wool Textile Organisation, 2000), re-

search typically focusses on binary classiﬁcation, e.g.

Rippel, O., Gülçelik, S., Rahimi, K., Kurniadi, J., Herrmann, A. and Merhof, D.

Animal Fiber Identiﬁcation under the Open Set Condition.

DOI: 10.5220/0010769800003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 5: VISAPP, pages

36-47

ISBN: 978-989-758-555-5; ISSN: 2184-4321

distinguishing between cashmere and wool (Robson,

1997; Robson, 2000; Xing et al., 2020a; Xing et al.,

2020b) or mohair and wool (Yildiz, 2020). (II) De-

veloped solutions lack the ability to reject ﬁber types

unseen during training at test time. Such a rejection

can be achieved in principle by out-of-distribution

(OOD)-detection techniques (Geng et al., 2020), fa-

cilitating the application of algorithms under the open

set condition (i.e. when training and test data do not

originate from the same data distribution). However,

the applicability of OOD-detection techniques to the

task at hand has not yet been demonstrated, and is the

main goal of our work.

Our contributions are as follows:

• We set up an exhaustive dataset comprising SEM-

images of 4 animal ﬁber types from 10 different

sources. In total, the dataset contains 6500 images

and covers the major axes of variation in natural

ﬁber surface morphology.

• We expand the open set F-measure

(Mendes J

unior et al., 2017) to facilitate the

evaluation of joint in-distribution classiﬁcation &

OOD-detection performance for OOD-detection

methods that output continuous OOD scores.

The proposed metric can be used to assess joint

performance across all possible OOD thresholds.

• We propose to perform OOD detection by mod-

eling the joint distribution of in-distribution data

across feature levels of converged, convolutional

neural network (CNN)-based classiﬁers via Gaus-

sian mixture models (GMMs). We compare this

approach to state-of-the-art OOD-detection algo-

rithms.

• We also investigate effects of outlier exposure

(OE) on animal ﬁber identiﬁcation under the open

set condition.

2 RELATED WORK

So far, the performance of animal ﬁber identiﬁca-

tion algorithms has not yet been investigated under

the open set condition. We will therefore give a

short deﬁnition of open set recognition (OSR) and re-

lated terminology ﬁrst. This will be followed by a

brief overview of proposed OOD-detection methods

as well as OE. Last, we will also present the open set

F-measure, which can be used to evaluate algorithms

under the open set condition.

2.1 Open Set Recognition

In general, OSR is concerned with problems that

arise speciﬁcally when training and test data do

not originate from the same data distribution (Geng

et al., 2020). In context of classiﬁcation, this means

one is ﬁrst tasked with distinguishing between in-

distribution and OOD data followed by the sub-

sequent d-class classiﬁcation of the presumed in-

distribution data. The in-distribution is furthermore

composed of the d target classes, which are re-

ferred to as known known classes (KKCs) (e.g. spe-

cialty ﬁbers such as cashmere). Opposed to the in-

distribution, the OOD is composed of known un-

known classes (KUCs), i.e. negative OOD samples

available during training/validation, and unknown un-

known classes (UUCs), i.e. samples not available for

training/validation.

The task of OOD detection can now be formulated

as “developing a measure which achieves lower val-

ues for in-distribution data compared to OOD data,

allowing for their separation”.

2.2 Methods for OOD Detection

OOD-detection methods commonly make use either

of the classiﬁer’s unnormalized predictions (also re-

ferred to as “logits” in literature) or of the underly-

ing feature representations to formulate their OOD

scores.

2.2.1 OOD Detection based on Unnormalized

Predictions

The most straightforward method is to take the maxi-

mum softmax probability (MSP) of the unnormalized

predictions as the OOD score (Hendrycks and Gim-

pel, 2017). Speciﬁcally, let φ : R

c×h×w

→ R

be a

pre-trained CNN classifying into d classes. Then, for

a given input x, the softmax probability s

for class i

with i ∈

{

1, . . . , d

}

is calculated as

(x) =

(x)

∑

(x)

. (1)

The probability that a given input sample belongs

to class i is given by s

(x). The MSP now takes

−max

(x) as the OOD score, since the classiﬁer

should be uncertain for samples not originating from

the in-distribution. Since it has been shown that clas-

siﬁers suffer from so-called overconﬁdent predictions

(Nguyen et al., 2015), i.e. from assigning high prob-

abilities to UUCs, modiﬁcations have been devel-

oped. For example, temperature scaling and input-

preprocessing have been proposed to improve MSP

Animal Fiber Identiﬁcation under the Open Set Condition

performance in outlier detection using in-degree num-

ber (ODIN) (Liang et al., 2018). Additionally, the

classiﬁer’s output has been used to deﬁne the energy-

based score (EBS)

EBS = − log

∑

(x)

(2)

based on the ties between energy-based modeling and

machine learning (Liu et al., 2020). Alternatively, it

has been proposed to use the maximum of the un-

normalized predictions (MaxLogit) as OOD score, ar-

guing that putting the class-predictions in relation to

each other via the softmax operator may be detrimen-

tal when semantically similar classes are present in

the in-distribution (Hendrycks et al., 2019a).

2.2.2 OOD Detection based on Intermediate

Features

Complementary to the classiﬁer’s predictions, the in-

termediate feature representations of a CNN can also

be used for OOD detection. Here, algorithms com-

monly try to estimate the probability density function

(PDF) of the in-distribution, often using the Gaus-

sian assumption (Rippel et al., 2021b; Kamoi and

Kobayashi, 2020) and its mixture models (Lee et al.,

2018b; Ahuja et al., 2019). Alternatively, deep gener-

ative models have also been used to ﬁt unconstrained

PDFs to in-distribution data in intermediate features

(Kirichenko et al., 2020; Zisselman and Tamar, 2020;

Zhang et al., 2020; Blum et al., 2021). The OOD

score is then deﬁned as the negative log-likelihood

(NLL) of a given input image x under the estimated

PDF.

Apart from the PDF-estimation, it has also been

proposed to use the distance of an input image x to

a ﬁxed reference UUC point inside the intermediate

features, resulting in the feature space singularity dis-

tance (FSSD) (Huang et al., 2020). Here, it is pro-

posed to use uniformly distributed noise samples as

the reference UUC point, arguing that uniform noise

possesses the highest degree of OOD-ness.

OOD detection has also been performed by au-

toencoders (AEs), where the encoder’s features are

simultaneously used to classify images into d KKCs

and to reconstruct the input images via the decoder

(Oza and Patel, 2019; Sun et al., 2020; Neal et al.,

2018). The OOD score is then deﬁned based on the

residual of the image reconstruction, which arguably

should be higher for OOD than in-distribution data.

It should be noted that these approaches roughly dou-

ble the computational complexity of the CNN and fail

to consistently achieve state-of-the-art results, and are

therefore not further regarded in this work.

2.3 Outlier Exposure

The OOD-detection methods from subsection 2.2 can

be applied to any converged CNN and require no

knowledge about OOD data. To now include avail-

able information about OOD data, it has been pro-

posed to use KUCs during training by means of OE

(Hendrycks et al., 2019c). In principle, an additional

loss term is introduced for the proposed OOD score

that is maximized for OOD samples (and optionally

minimized for in-distribution samples), e.g. Equa-

tion 2. While it has been shown that sampling the

OOD introduces a bias in the resulting OOD detector,

i.e. OE may actually hurt OOD detection for some

UUCs (Ye et al., 2021), OE has also been shown to

work for most UUCs (Hendrycks et al., 2019c). As an

alternative to real OOD images, approaches facilitat-

ing OE by means of synthetic OOD images have also

been proposed (Neal et al., 2018; Lee et al., 2018a;

Grci

c. et al., 2021).

2.4 Open Set F-measure

In order to jointly assess OOD-detection and in-

distribution classiﬁcation performance, the open set

F-measure has been proposed (Mendes J

unior et al.,

2017). Speciﬁcally, for a d-class classiﬁcation prob-

lem, the open set F-measure for class i is deﬁned as

the harmonic mean between its open set precision and

recall, deﬁned as

T P

+ FP

, FP

∑

j=1

i, j

+ FP

i,UUC

(3)

and

T P

+ FN

, FN

∑

j=1

i, j

+ FN

i,UUC

(4)

respectively. Compared to the closed set precision

and recall variants, it can be seen that false posi-

tives may now also be incurred by failing to reject an

OOD image followed by its misclassiﬁcation to class i

Equation 3, and false negatives may be incurred by in-

correctly labeling in-distribution data as OOD Equa-

tion 4 (refer also Figure 1a).

As can be inferred from Equation 3 and Equa-

tion 4, the open set F-measure requires that every im-

age is labeled either as OOD and rejected or labeled as

in-distribution & subsequently classiﬁed. It is there-

fore ill-suited for OOD-detection methods that yield

continuous OOD scores and thus require thresholding

to achieve the aforementioned partitioning of images

into in-distribution and OOD.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

3 NATURAL FIBER

IDENTIFICATION UNDER THE

OPEN SET CONDITION

We specify both our proposed OOD-detection method

as well as the threshold-independent evaluation of

joint in-distribution classiﬁcation & OOD-detection

performance in the following.

3.1 OOD Detection via Modeling the

Joint PDF Across Feature Levels

Similar to (Ahuja et al., 2019), we model the PDF

of in-distribution data of a converged CNN by means

of GMMs. However, as a signiﬁcant extension, we

propose to model the joint PDF of in-distribution data

across feature levels instead of modeling PDFs layer-

wise and summing their individual OOD scores. We

motivate this by the fact that consistency of an in-

put image x across feature levels of a network has

been shown to be an indicator of model generaliza-

tion, i.e. models with consistent representations gen-

eralize well (Natekar and Sharma, 2020). We argue

that inconsistent representations may be an indicator

of OOD, and show that joint PDF-estimation is ben-

eﬁcial for OOD detection empirically in subsubsec-

tion 4.3.2.

Speciﬁcally, let φ : R

c×h×w

→ R

again be a pre-

trained CNN classifying into d classes. Each inter-

mediate layer’s output ψ

:= φ

(ψ

m−1

) for an input

image ψ

:= x

has c

features and spatial dimension

by w

. We now reduce over the spatial dimen-

sions h

and w

by means of averaging, resulting in

a feature vector c extracted per intermediate layer m.

By concatenating c of all intermediate layers, a larger

feature vector c

cat

is generated. To model the PDF of

in-distribution data in c

cat

, we make use of GMMs,

deﬁned as

p(x) =

∑

i=1

N (x|µ

, Σ

), (5)

with

∑

i=1

= 1, k being the number of Gaussian mix-

ture components and µ

and Σ

denoting the mean

vector and covariance matrix of mixture component

i. We approximate the parameters of the GMM by

using the expectation maximization (EM) algorithm,

and determine the number of Gaussians k by means

of the Bayesian information criterion (BIC) (Bishop,

2006). Similar to other PDF-estimation approaches,

we use the NLL of an input image x under the esti-

mated PDF as the OOD score.

3.2 Threshold-independent Evaluation

of Joint In-distribution

Classiﬁcation & OOD-detection

Performance

When looking at recent work on OOD detection,

it becomes apparent that OOD-detection and in-

distribution classiﬁcation performances are reported

individually (Liu et al., 2020; Lee et al., 2018b;

Hendrycks and Gimpel, 2017; Liang et al., 2018).

We argue that reporting OOD-detection and in-

distribution classiﬁcation performances separately

oversimpliﬁes the open set classiﬁcation problem,

and verify this claim experimentally in subsubsec-

tion 4.3.1.

Since most OOD-detection approaches proposed

in literature yield continous OOD scores, we extend

the open set F-measure (refer subsection 2.4) to be

threshold-independent. Speciﬁcally, we propose to

generate open set precision-recall curves (Davis and

Goadrich, 2006) by plotting the open set precision

& recall values (Equation 3 and Equation 4) across

all potential OOD thresholds t (refer Figure 1). Such

curves are generated for each class individually, and

the area under the open set precision-recall curve

(AUPR

) can now be used to measure the joint in-

distribution classiﬁcation & OOD-detection perfor-

mance of a single class.

As we want to quantify the overall performance

across all classes, we furthermore propose to com-

pute the average of the class-wise open set preci-

sion & recall values for each threshold t. The re-

sulting values can then again be plotted for all t

to yield a class-averaged, open set precision-recall

curve (refer the yellow curve in Figure 1c). The area

under the class-averaged, open set precision-recall

curve (AUPR

os,avg

) can now be used to jointly as-

sess the overall in-distribution classiﬁcation & OOD-

detection performance for multi-class classiﬁcation

problems. Similar to the macro-averaged F1-score,

AUPR

os,avg

regards all KKCs as equally important.

Since not all OOD-detection methods provide contin-

uous OOD scores, we also report the Euclidean dis-

tance of the class-averaged, open set precision-recall

curve to the optimal point (1, 1), yielding PR

dist

for

comparison.

4 EXPERIMENTS

In the following, we conduct experiments to assess

the applicability of natural ﬁber identiﬁcation algo-

rithms under the open set condition.

Animal Fiber Identiﬁcation under the Open Set Condition

OOD

Predicted label

True label

OOD

OOD score

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

avg,opt

, P

avg,opt

)

Recall

Precision

os,avg

(a)

(b)

(c)

Figure 1: Threshold-independent evaluation of joint OOD-detection and in-distribution classiﬁcation performance for multi-

class classiﬁcation problems. The open set confusion matrix at a single OOD threshold t is shown in (a), whereas (b) shows

a scatter plot of OOD scores for a hypothetical 3-class open set classiﬁer. By iterating over all possible OOD thresholds t,

class-wise as well as a class-averaged open set precision-recall curves can be computed, shown in (c). The areas under the

open set precision-recall curves can now be used to assess joint OOD-detection and in-distribution classiﬁcation performance.

4.1 Dataset

We set up an exhaustive dataset to facilitate the per-

formed experiments. We employ SEM-imaging over

optical microscopy as it provides higher-resolution

images of the ﬁber surface morphology, and is the

only imaging technique shown to be reliable in com-

bination with human operators (Wortmann and Wort-

mann, 1992). Our dataset contains 4 animal ﬁber

types from 10 different sources, and 6500 images

in total. Furthermore, all specimen samples were

checked for purity & identity by a certiﬁed laboratory

prior to image acquisition.

When composing the dataset, focus was put on

sampling the three major axes of variation in natural

ﬁber surface morphology:

1. Inter-species variation. These are variations

present between different species, e.g. cashmere

& yak.

2. Intra-species variation. These are variations

present between different races of the same

species, e.g. merino wool and typical wool.

3. Treatment status variation. These are variations

introduced by the mechanical & chemical pro-

cesses applied to the ﬁbers during fabric produc-

tion (e.g. they are dyed & bleached (d&b)). Vari-

ations incurred by treatment status are important

for the industrial application of animal ﬁber iden-

tiﬁcation since adulteration attempts often involve

treatment of non-specialty ﬁbers to make their

surface morphology more similar to the one of

specialty ﬁbers.

Reference images for all three axes of variation

are shown in Figure 2.

Table 1: Characteristics of the animal ﬁber dataset used in

this work.

Fiber Type #Samples Use

Cashmere, Iranian 500 KKC

Cashmere, Chinese 500 KKC

Cashmere, brown 500 KKC

Yak, type 1 500 KKC

Yak, type 2 500 KKC

Wool, Suedwolle 1000 KKC

Wool, Interwool 1000 KKC

Wool, Suedwolle d&b 1000 UUC/KUC

Wool, merino 500 UUC/KUC

Silk 500 UUC/KUC

4.2 Experimental Setup

In the following, we describe the experimental setup

required to assess the applicability of natural ﬁber

identiﬁcation algorithms under the open set condition.

4.2.1 Dataset Composition & Splits

In all experiments, the KKCs are cashmere, yak and

wool (refer Table 1). We pool all sources for the three

KKCs as it is not necessary to distinguish between

subtypes of animal ﬁbers under current regulations

(Council of European Union, 2011; Freer, 1946). In

order to represent all three possible axes of varia-

tion the UUCs/KUCs are: silk (inter-species), merino

wool (intra-species) and d&b wool (treatment status).

Since all axes are equally important, we will report

the evaluation scores per axis and give their mean and

standard deviation to denote overall algorithmic per-

formance.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

(a) Cashmere, Chinese (b) Yak, type 1

(e) Wool, merino (f) Silk

Figure 2: Reference sample images from the animal ﬁber dataset.

Following good scientiﬁc practice, we split our

dataset into training, validation and test splits. Since

KKCs, UUCs and KUCs serve different purposes dur-

ing model training, validation and testing, they are

split independently.

The KKCs are used to train the in-distribution

classiﬁer for classifying into d = 3 classes. Since the

dataset is rather small, we perform a 5-fold cross-

validation to improve the robustness of our evalua-

tions. To this end, the data is split in a 3-1-1 split,

meaning that in each fold 60% of the data is used for

training, and 20% is used each for validation and test-

ing. Splits are furthermore stratiﬁed according to the

prevalence of the KKCs.

Assessing algorithmic performance under the

open set condition requires UUCs. Since UUCs

are per-deﬁnition unknown during model train-

ing/evaluation, they are only used at test time. As

is common in literature, the size of the UUC dataset

is set to a ﬁfth of the KKCs’ testing set (Hendrycks

et al., 2019c; Liu et al., 2020). In order to cover most

of the variation of the UUC data, the model is tested

with a randomly sampled UUC set for each fold.

For the experiments on OE (subsubsection 4.3.3),

KUCs are also used during training. Identical to the

UUCs, KUCs also compose a ﬁfth of the KKCs’ train-

ing set and are also sampled randomly for each fold.

In addition to the performance on the KUC class, we

will also report the OOD-detection performance on

the two remaining UUCs to assess whether potential

gains on the KUC also generalize to the UUCs.

4.2.2 Evaluation Details

Apart from the proposed joint-performance metric

(subsection 3.2), we also report OOD-detection per-

formance and in-distribution classiﬁcation perfor-

mance individually. For the OOD-detection per-

formance, we compute the area under the receiver

operating characteristic (ROC) curve (AUROC) of

the binary in-distribution/OOD classiﬁcation prob-

lem, where in-distribution data is the positive and

OOD data is the negative class. For the in-distribution

classiﬁcation performance, we compute the macro-

averaged F1-score of the in-distribution test set.

4.2.3 Model Architecture & Evaluated

OOD-detection Methods

For all our experiments, we employ the EfﬁcientNet-

B0 (Tan and Le, 2019) model architecture, which con-

sists of nine levels. Furthermore, we perform trans-

fer learning with an ImageNet(Deng et al., 2009)-

pretrained EfﬁcientNet-B0, as pre-training on large-

Animal Fiber Identiﬁcation under the Open Set Condition

(a) (b)

Figure 3: Iranian cashmere sample before (a) and after (b)

the application of CLAHE.

scale datasets improves model robustness and uncer-

tainty (Hendrycks et al., 2019b). For our proposed

method, we extract features from levels 2, 3, 4, 6 and

We compare our proposed method with the Ma-

halanobis (Maha) score (Lee et al., 2018b), MSP,

ODIN, EBS, MaxLogit and the FSSD. Note that for

Maha, we extract features from the same levels as for

our approach, and give the unweighted mean of level-

wise scores as the overall OOD score.

4.2.4 Image Preprocessing & Training Details

To compensate for inhomogeneities in image contrast

caused by the image acquisition procedure, contrast

limited adaptive histogram equalization (CLAHE)

(Pizer et al., 1987) is applied (refer Figure 3 for a ref-

erence image). Further, the images are normalized

and scaled down to the expected input resolution of

the pre-trained classiﬁer (224 ×224 pixels).

We ﬁne-tune the pre-trained EfﬁcientNet-B0 us-

ing the cross-entropy loss (Goodfellow et al., 2016)

in combination with the Adam optimizer (Kingma

and Ba, 2015), a learning rate of 0.0001 and batch-

size of 16. The training set is used to train the 3-

class classiﬁer, and the validation set is used to de-

tect the best model by calculating the macro-averaged

F1-score over in-distribution data only. Model train-

ing is furthermore stopped when no improvement of

at least 2% is achieved within 10 epochs for the in-

distribution F1-score. Moreover, the validation set is

used to parametrize the OOD-detection methods after

the model has converged (i.e. estimate the joint in-

distribution PDF by means of GMM). The test set is

subsequently used to test model performance.

For experiments that employ OE, we use samples

from the KUC to minimize

(φ(x

KUC

)) =log

∑

i=1

KUC

)

−

∑

i=1

KUC

)

(6)

in addition to the supervised cross-entropy loss,

which is similar to (Hendrycks et al., 2019c). The

Fold F-score

0.0 96.2

1.0 95.5

2.0 96.4

3.0 95.3

4.0 96.4

µ 96.0

σ 0.5

(a)

Y W

0.95 0.05 0.00

0.05 0.94 0.01

0.02 0.00 0.98

(b)

Figure 4: In-distribution classiﬁcation performance without

application of OE. (a) shows per-fold F1-scores whereas (b)

shows the corresponding confusion matrix of fold 1.

overall loss for training with OE is thus given as

L = L

(φ(x

KKC

)) +λL

(φ(x

KUC

)). (7)

Based on preliminary experiments, we set λ to 0.5.

4.3 Results

We ﬁrst compare with state-of-the-art OOD-detection

methods in subsubsection 4.3.1. Afterwards, we per-

form an ablation study to investigate the inﬂuence of

the proposed PDF-modeling across feature levels on

joint performance in subsubsection 4.3.2. Next, we

assess inﬂuence of OE on open set classiﬁcation per-

formance in subsubsection 4.3.3.

4.3.1 Natural Fiber Identiﬁcation Performance

under the Open Set

Table 2 shows that joint performance of the individual

algorithms varies with the axes of biological variation

that needs to be detected as OOD. Here, it can be seen

that our proposed method is best-suited for detecting

modiﬁcations incurred by chemical treatments (wool

d&b), making it especially suitable for detecting adul-

teration attempts. Furthermore, our proposed method

performs best and most consistent over all axes of

variation, achieving an AUPR

os,avg

of 91.9 ± 2.1.

Regarding in-distribution classiﬁcation perfor-

mance, it can be seen that classiﬁcation rates com-

parable to human raters are achieved with a macro-

averaged in-distribution F1-score of 96.0 ± 0.5. Note

that all methods share the same F1-scores given in

Figure 4a.

Assessing pure OOD-detection performance, it

can be seen that our method again performs the

most consistent across all axes of biological variation,

achieving an AUROC of 84.8 ± 8.4 (Table 3). More-

over, Maha surprisingly performs best with respect

to OOD detection for two of the three UUCs while

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

Table 2: AUPR

os,avg

(%) for different UUCs. Best value per row is boldfaced.

prediction-based feature-based

UUC MSP ODIN EBS MaxLogit Maha FSSD Ours

Wool d&b 86.6 85.5 85.5 85.5 90.5 79.9 93.8

Merino 93.0 93.0 93.0 93.1 74.5 74.3 89.7

Silk 93.5 92.2 91.6 91.8 90.3 80.7 92.1

µ 91.0 90.3 90.0 90.1 85.1 78.3 91.9

σ 3.9 4.1 4.0 4.1 9.2 3.5 2.1

Table 3: AUROC (%) for different UUCs. Best value per row is boldfaced.

prediction-based feature-based

UUC MSP ODIN EBS MaxLogit Maha FSSD Ours

Wool d&b 32.4 31.6 31.8 31.8 94.7 35.2 92.8

Merino 82.6 83.4 84.1 84.0 38.6 32.4 76.0

Silk 89.9 84.9 83.4 84.1 96.2 59.1 85.5

µ 68.3 66.6 66.4 66.6 76.5 42.2 84.8

σ 31.3 30.4 30.0 30.2 32.8 14.7 8.4

achieving subpar AUPR

os,avg

values for them (refer

Table 2).

We investigate the reasons behind this next by as-

sessing the class-averaged, open set precision-recall

curve for both Maha and our method on the UUC silk

in Figure 5. Here, it can be seen that the curve of

our proposed method possesses desirable characteris-

tics, as no major dips in precision can be observed for

low recall values, indicating that correctly classiﬁed

in-distribution samples of all classes achieve lowest

OOD scores. Conversely, Maha achieves higher pre-

cision & recall values later on, but exhibits dips in

precision for low recall values. In combination with

high OOD-detection performance, this indicates that

lowest OOD scores are assigned to misclassiﬁed in-

distribution data. Therefore, it is important to as-

sess the joint-performance of OOD-detection & in-

distribution classiﬁcation when evaluating classiﬁers

under the open set condition.

4.3.2 Ablation Study

We perform an ablation experiment to assess the im-

portance of modeling the joint PDF of in-distribution

data across feature levels of a classiﬁer. To this end,

we compare our approach to ﬁtting GMMs to the in-

distribution data in every feature level individually,

giving their average NLL as the overall OOD score.

Assessing results in Table 4, it can be seen that

the proposed joint-modeling boosts OOD-detection

performance measured by AUROC as well as joint

performance as measured by AUPR

os,avg

for all eval-

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Recall

Precision

Ours

Maha

Figure 5: PR

os,avg

curve of the proposed OOD-method as

well as Maha. UUC is silk.

uated UUCs. While lower standard deviations are

achieved for the level-wise PDF-estimation, this is ex-

plained by the fact that the improvement of joint PDF-

estimation varies with respect to the chosen UUC

class.

4.3.3 Inﬂuence of OE

We assess the inﬂuence of OE on natural ﬁber identi-

ﬁcation under the open set condition next. To this end,

we iterate over the three UUCs, using them as KUC

for OE. In addition to OE, we also evaluate a simple

k+1 classiﬁer trained with the KUC as the reject class.

The results in Table 5 show that OE on average

decreases joint performance, as indicated by larger

dist

-values for all methods. In fact, OE does im-

Animal Fiber Identiﬁcation under the Open Set Condition

Table 4: AUROC (%) and AUPR

os,avg

(%) for joint

PDF-estimation across feature levels vs. level-wise PDF-

estimation followed by averaging of level-wise OOD

scores. Best values per metric and row are boldfaced.

UUC

AUROC AUPR

os,avg

level-wise joint level-wise joint

Wool d&b 91.9 92.8 93.3 93.8

Merino 75.4 76.0 89.5 89.7

Silk 83.6 85.5 91.6 92.1

µ 83.6 84.8 91.5 91.9

σ 8.3 8.4 1.9 2.1

Class µ σ

Wool d&b 93.0 1.4

Merino 95.5 1.0

Silk 92.6 1.7

no OE 96.0 0.5

(a)

Y W

0.84 0.16 0.00

0.09 0.90 0.01

0.01 0.02 0.96

(b)

Figure 6: In-distribution classiﬁcation performance under

application of OE. (a) shows per-fold F1-scores whereas (b)

shows the corresponding confusion matrix of fold 1 with

KUC = Wool d&b.

prove joint performance only for the KUC, and sig-

niﬁcantly reduces the performance for the UUCs not

used for OE. The aforementioned effect is further-

more strongest for the k+1 classiﬁer, and its inverse

can be observed for FSSD.

When investigating the mechanisms underlying

this phenomenon by assessing OOD detection and

in-distribution classiﬁcation individually, it can be

seen that OE does improve OOD-detection perfor-

mance on average (Table 6) at the cost of a reduced

in-distribution classiﬁcation performance (Figure 6a).

Moreover, the OOD-detection results improve sub-

stantially only for the KUC class. In fact, they ac-

tually degrade for the UUCs across all assessed al-

gorithms with the exception of Maha. Thus, OOD-

detection improvements achieved by OE for KUCs

do not propagate to UUCs. Coupled with an over-

all decrease in in-distribution classiﬁcation perfor-

mance, this translates to overall beneﬁts as measured

by PR

dist

only for the KUCs.

5 DISCUSSION

In our work, we have investigated the performance of

natural ﬁber identiﬁcation algorithms under the open

set condition. To this end, we identiﬁed the three

main axes of variation that classiﬁers need to be ro-

bust against & set up a dataset that reﬂects these vari-

ations.

Experiments revealed that the proposed joint

PDF-modeling across feature levels of a CNN per-

forms best overall. Moreover, the obtained in-

distribution classiﬁcation rates were high enough to

warrant a potential model deployment (>96% agree-

ment with the nominal value is required for human

operators (Zhang and Ainsworth, 2005)). The ab-

lation study in subsubsection 4.3.2 further showed

that the joint PDF-modeling was beneﬁcial for all

UUCs. This shows that consistency of model rep-

resentations is not only predictive of model general-

ization (Natekar and Sharma, 2020), but can further-

more be used to boost OOD-detection performance of

feature-based OOD-detection methods. Note that the

joint modeling across feature levels was also shown to

be beneﬁcial for transfer-learning anomaly detection

(AD), which is concerned with performing OOD de-

tection under the one-class-classiﬁcation setting (De-

fard et al., 2020).

Moreover, the importance of assessing joint OOD-

detection & in-distribution classiﬁcation performance

became evident in subsubsection 4.3.1, where Maha

achieved subpar joint performance as measured by

AUPR

os,avg

despite achieving strong OOD-detection

and in-distribution classiﬁcation results. Therefore,

we argue that one should report joint performance in

future when evaluating algorithms under the open set

condition rather than assessing in-distribution perfor-

mance and OOD-detection performance separately, as

was best practice so far (Liu et al., 2020; Lee et al.,

2018b; Hendrycks and Gimpel, 2017; Liang et al.,

2018). Here, the next step is to expand the proposed

AUPR

os,avg

to the binary classiﬁcation as well as to

the object detection task.

We also assessed the inﬂuence of OE on natural

ﬁber identiﬁcation under the open set condition. In-

terestingly, it was found that performance improve-

ments for the KUC used for OE did not propagate to

the UUCs. These results are in line with the ﬁnd-

ing that a bias is introduced by sampling OOD data

for OE, which may negatively impact OOD-detection

performance of UUCs(Ye et al., 2021). Furthermore,

OE reduced in-distribution classiﬁcation performance

in our experiments. Therefore, natural ﬁber identiﬁ-

cation under the open set condition does not beneﬁt

from OE, especially since many UUCs, e.g. adulter-

ation procedures, are present.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

Table 5: PR

dist

(%) for open set classiﬁcation with OE. µ and σ are calculated over all possible UUC – KUC combinations

(All), over all combinations where UUC ̸= KUC, over all combinations where UUC = KUC and, for comparison, when no

OE is applied.

prediction-based feature-based

MSP ODIN EBS MaxLogit Maha FSSD Ours k+1

All

µ 17.1 17.9 17.9 17.6 15.5 20.0 15.6 14.7

σ 2.1 2.6 3.1 2.9 4.3 3.9 3.6 6.5

UUC = KUC

µ 16.5 17.3 17.0 16.9 15.2 21.4 13.8 6.5

σ 2.4 3.2 3.4 3.2 3.3 1.8 3.5 0.3

UUC ̸= KUC

µ 17.3 18.1 18.3 18.0 15.6 19.2 16.5 18.8

σ 1.8 2.5 2.7 2.6 4.8 4.5 3.6 2.8

No OE

µ 14.5 15.7 16.2 15.9 13.8 17.9 15.0 —

σ 0.5 1.9 3.0 2.6 4.7 3.1 2.9 —

Table 6: AUROC (%) for OOD-detection with OE. µ and σ are calculated over all possible UUC – KUC combinations (All),

over all combinations where UUC ̸= KUC, over all combinations where UUC = KUC and, for comparison, when no OE is

applied.

prediction-based feature-based

MSP ODIN EBS MaxLogit Maha FSSD Ours

All

µ 70.7 69.9 69.8 70.4 82.5 43.5 87.1

σ 21.4 20.7 20.9 20.8 24.3 21.8 13.3

UUC = KUC

µ 79.0 80.5 82.2 81.7 90.8 50.1 94.0

σ 17.8 14.9 14.3 14.9 8.7 25.2 5.5

UUC ̸= KUC

µ 66.5 64.6 63.7 64.7 78.4 40.2 83.7

σ 22.0 21.2 21.0 21.1 29.0 18.4 15.1

No OE

µ 68.3 66.6 66.4 66.6 76.5 42.2 84.8

σ 31.3 30.4 30.0 30.2 32.8 14.7 8.4

5.1 Limitations

While, to the best of our knowledge, we have used

the largest and most diverse dataset so far, human op-

erators are capable of distinguishing between at least

11 ﬁber types (10 specialty ﬁbers + wool) (Interna-

tional Wool Textile Organisation, 2000). We will

therefore expand the dataset, focussing on covering as

many ﬁber types from as diverse sources as possible.

Moreover, while our proposed method performed best

overall, it did not achieve highest values for every sin-

gle UUC. In fact, the best performance for two of the

three UUCs was achieved by methods that are based

on the classiﬁer’s unnormalized predictions (refer Ta-

ble 2). In our future work, we will therefore develop

OOD-detection methods that leverage both interme-

diate feature representations as well as the classiﬁer’s

output. Last, we did not assess the performance of the

model when distribution shifts occur for the KKCs.

Ideally, the classiﬁer would be robust to this under

the open set condition, i.e. it would accept & cor-

rectly classify KKCs which have undergone input-

distribution shifts rather than rejecting them. This

would require the OOD-detection method to assign

lower OOD scores to KKCs that have undergone dis-

tribution shifts compared to UUCs. Advances from

the ﬁeld of domain adaptation can be used as a start-

ing point here (Saito and Saenko, 2021; Bashkirova

et al., 2021).

6 CONCLUSION

In our work, we have thoroughly investigated the per-

formance of natural ﬁber identiﬁcation algorithms un-

der the open set condition. To this end, we identiﬁed

the three main axes of variation that natural ﬁber iden-

tiﬁcation algorithms need to be robust against, and

have created a dataset that is able to reﬂect them. Our

experiments revealed that the proposed joint PDF-

modeling across feature levels of a CNN performs

best, achieving highest AUPR

os,avg

amongst all eval-

uated methods. Furthermore, we demonstrated that

metrics of joint performance are necessary to fully re-

Animal Fiber Identiﬁcation under the Open Set Condition

capitulate the behavior of a classiﬁer under the open

set condition. Our work thus shows that natural ﬁber

identiﬁcation algorithms provide promising results in

real-world scenarios, i.e. under the open set condition.

Our future work will focus on improving the OOD-

detection performance and investigating the behav-

ior of the classiﬁer when simultaneously challenged

with distribution shifts of KKCs and OOD-detection

of UUCs.

ACKNOWLEDGEMENTS

This work was supported by the German Federation

of Industrial Research Associations (AiF) under the

grant number 21376 N.

REFERENCES

Ahuja, N. A., Ndiour, I., Kalyanpur, T., and Tickoo, O.

(2019). Probabilistic modeling of deep features for

out-of-distribution and adversarial detection. arXiv

preprint arXiv:1909.11786.

American Society for Testing and Materials (1993). Stan-

dard methods for quantitative analysis of textiles,

method d629-88.

Bashkirova, D., Hendrycks, D., Kim, D., Mishra, S.,

Saenko, K., Saito, K., Teterwak, P., and Usman,

B. (2021). Visda-2021 competition universal do-

main adaptation to improve performance on out-of-

distribution data. arXiv preprint arXiv:2107.11011.

Bishop, C. (2006). Pattern Recognition and Machine

Learning. Springer.

Blum, H., Sarlin, P.-E., Nieto, J., Siegwart, R., and Cadena,

C. (2021). The ﬁshyscapes benchmark: Measuring

blind spots in semantic segmentation. International

Journal of Computer Vision, 129(11):3119–3135.

Council of European Union (2011). Council regulation

(EU) no 1007/2011.

Davis, J. and Goadrich, M. (2006). The relationship be-

tween precision-recall and roc curves. In Proceed-

ings of the 23rd international conference on Machine

learning, pages 233–240.

Defard, T., Setkov, A., Loesch, A., and Audigier, R. (2020).

Padim: a patch distribution modeling framework for

anomaly detection and localization. arXiv preprint

arXiv:2011.08785.

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

L. (2009). ImageNet: A large-scale hierarchical im-

age database. In 2009 IEEE Conference on Computer

Vision and Pattern Recognition, pages 248–255.

Freer, R. E. (1946). The wool products labeling act of 1939.

Temp. LQ, 20:42.

Geng, C., Huang, S.-J., and Chen, S. (2020). Recent ad-

vances in open set recognition: A survey. IEEE Trans-

actions on Pattern Analysis and Machine Intelligence,

pages 1–1.

Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep

learning. MIT press.

Grci

c., M., Bevandi

c., P., and Segvi

c., S. (2021). Dense

open-set recognition with synthetic outliers generated

by real nvp. In Proceedings of the 16th Interna-

tional Joint Conference on Computer Vision, Imag-

ing and Computer Graphics Theory and Applications

- Volume 4: VISAPP,, pages 133–143. INSTICC,

SciTePress.

Hendrycks, D., Basart, S., Mazeika, M., Mostajabi, M.,

Steinhardt, J., and Song, D. (2019a). Scaling out-of-

distribution detection for real-world settings. arXiv

preprint arXiv:1911.11132.

Hendrycks, D. and Gimpel, K. (2017). A baseline for de-

tecting misclassiﬁed and out-of-distribution examples

in neural networks. In 5th International Conference

on Learning Representations, ICLR 2017, Toulon,

France, April 24-26, 2017, Conference Track Pro-

ceedings. OpenReview.net.

Hendrycks, D., Lee, K., and Mazeika, M. (2019b). Us-

ing pre-training can improve model robustness and

uncertainty. In International Conference on Machine

Learning, pages 2712–2721. PMLR.

Hendrycks, D., Mazeika, M., and Dietterich, T. (2019c).

Deep anomaly detection with outlier exposure. In In-

ternational Conference on Learning Representations.

Huang, H., Li, Z., Wang, L., Chen, S., Dong, B.,

and Zhou, X. (2020). Feature space singular-

ity for out-of-distribution detection. arXiv preprint

arXiv:2011.14654.

International Wool Textile Organisation (2000). Test

method no. iwto-58-00: Scanning electron micro-

scopic analysis of speciality ﬁbres and sheep’s wool

and their blends.

International Wool Textile Organisation (2018). Statistics

for the global wool production and textile industry,.

Kamoi, R. and Kobayashi, K. (2020). Why is the ma-

halanobis distance effective for anomaly detection?

arXiv preprint arXiv:2003.00402.

Kim, Y., Kim, T., and Choi, H.-M. (2013). Qualita-

tive identiﬁcation of cashmere and yak ﬁbers by pro-

tein ﬁngerprint analysis using matrix-assisted laser

desorption/ionization time-of-ﬂight mass spectrome-

try. Industrial & Engineering Chemistry Research,

52(16):5563–5571.

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

stochastic optimization. In Bengio, Y. and LeCun,

Y., editors, 3rd International Conference on Learn-

ing Representations, ICLR 2015, San Diego, CA, USA,

May 7-9, 2015, Conference Track Proceedings.

Kirichenko, P., Izmailov, P., and Wilson, A. G. (2020). Why

normalizing ﬂows fail to detect out-of-distribution

data. In Larochelle, H., Ranzato, M., Hadsell, R.,

Balcan, M. F., and Lin, H., editors, Advances in Neu-

ral Information Processing Systems, volume 33, pages

20578–20589. Curran Associates, Inc.

Lee, K., Lee, H., Lee, K., and Shin, J. (2018a). Training

conﬁdence-calibrated classiﬁers for detecting out-of-

distribution samples. In International Conference on

Learning Representations.

Lee, K., Lee, K., Lee, H., and Shin, J. (2018b). A simple

uniﬁed framework for detecting out-of-distribution

samples and adversarial attacks. In Advances in

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

Neural Information Processing Systems, pages 7167–

7177.

Liang, S., Li, Y., and Srikant, R. (2018). Enhancing the reli-

ability of out-of-distribution image detection in neural

networks. In International Conference on Learning

Representations.

Liu, W., Wang, X., Owens, J., and Li, S. Y. (2020). Energy-

based out-of-distribution detection. Advances in Neu-

ral Information Processing Systems, 33.

McGregor, B. A. (2018). 4 - physical, chemical, and tensile

properties of cashmere, mohair, alpaca, and other rare

animal ﬁbers. In Bunsell, A. R., editor, Handbook

of Properties of Textile and Technical Fibres (Second

Edition), The Textile Institute Book Series, pages 105

– 136. Woodhead Publishing, second edition edition.

Mendes J

unior, P. R., de Souza, R. M., Werneck, R. d. O.,

Stein, B. V., Pazinato, D. V., de Almeida, W. R., Pe-

natti, O. A. B., Torres, R. d. S., and Rocha, A. (2017).

Nearest neighbors distance ratio open-set classiﬁer.

Machine Learning, 106(3):359–386.

Natekar, P. and Sharma, M. (2020). Representation based

complexity measures for predicting generalization in

deep learning. arXiv preprint arXiv:2012.02775.

Neal, L., Olson, M., Fern, X., Wong, W.-K., and Li, F.

(2018). Open set learning with counterfactual images.

In Proceedings of the European Conference on Com-

puter Vision (ECCV).

Nguyen, A., Yosinski, J., and Clune, J. (2015). Deep neural

networks are easily fooled: High conﬁdence predic-

tions for unrecognizable images. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 427–436.

Oza, P. and Patel, V. M. (2019). C2ae: Class conditioned

auto-encoder for open-set recognition. In Proceedings

of the IEEE/CVF Conference on Computer Vision and

Pattern Recognition (CVPR).

Phan, K.-H. and Wortmann, F. (2001). Quality assessment

of goat hair for textile use. In Silk, Mohair, Cashmere

and Other Luxury Fibres, pages 227–233. Elsevier.

Pizer, S. M., Amburn, E. P., Austin, J. D., Cromartie, R.,

Geselowitz, A., Greer, T., ter Haar Romeny, B., Zim-

merman, J. B., and Zuiderveld, K. (1987). Adaptive

histogram equalization and its variations. Computer

vision, graphics, and image processing, 39(3):355–

368.

Rane, P. P. and Barve, S. (2011). Standardization and op-

timization of mtdna isolation and molecular genetic

analysis of d-loop region in animal natural ﬁbres. In-

ternational Journal of Zoological Research, 7(2):190.

Rippel, O., Bilitewski, N., Rahimi, K., Kurniadi, J., Her-

rmann, A., and Merhof, D. (2021a). Identifying pris-

tine and processed animal ﬁbers using machine learn-

ing. In 2021 IEEE International Instrumentation and

Measurement Technology Conference (I2MTC), pages

1–6.

Rippel, O., Mertens, P., and Merhof, D. (2021b). Model-

ing the distribution of normal data in pre-trained deep

features for anomaly detection. In 2020 25th Inter-

national Conference on Pattern Recognition (ICPR),

pages 6726–6733.

Robson, D. (1997). Animal ﬁber analysis using imaging

techniques: Part i: Scale pattern data. Textile Research

Journal, 67(10):747–752.

Robson, D. (2000). Animal ﬁber analysis using imaging

techniques: Part ii: Addition of scale height data. Tex-

tile Research Journal, 70(2):116–120.

Saito, K. and Saenko, K. (2021). Ovanet: One-vs-all net-

work for universal domain adaptation. arXiv preprint

arXiv:2104.03344.

Sun, X., Yang, Z., Zhang, C., Ling, K.-V., and Peng, G.

(2020). Conditional gaussian distribution learning for

open set recognition. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion (CVPR).

Tan, M. and Le, Q. V. (2019). EfﬁcientNet: Rethinking

model scaling for convolutional neural networks. In

Chaudhuri, K. and Salakhutdinov, R., editors, Pro-

ceedings of the 36th International Conference on Ma-

chine Learning, ICML 2019, 9-15 June 2019, Long

Beach, California, USA, volume 97 of Proceedings

of Machine Learning Research, pages 6105–6114.

PMLR.

Waldron, S., Brown, C., and Komarek, A. M. (2014). The

chinese cashmere industry: a global value chain anal-

ysis. Development Policy Review, 32(5):589–610.

Wortmann, F.-J. (1991). Quantitative ﬁber mixture analy-

sis by scanning electron microscopy: Part iii: Round

trial results on mohair / wool blends. Textile Research

Journal, 61(7):371–374.

Wortmann, F.-J. and Wortmann, G. (1992). Quantita-

tive ﬁber mixture analysis by scanning electron mi-

croscopy: Part iv: Assessment of light microscopy as

an alternative tool for analyzing wool/specialty ﬁber

blends. Textile Research Journal, 62(7):423–431.

Xing, W., Liu, Y., Deng, N., Xin, B., Wang, W., and Chen,

Y. (2020a). Automatic identiﬁcation of cashmere and

wool ﬁbers based on the morphological features anal-

ysis. Micron, 128:102768.

Xing, W., Liu, Y., Xin, B., Zang, L., and Deng, N. (2020b).

The application of deep and transfer learning for iden-

tifying cashmere and wool ﬁbers. Journal of Natural

Fibers, 0(0):1–17.

Ye, Z., Chen, Y., and Zheng, H. (2021). Understanding

the effect of bias in deep anomaly detection. arXiv

preprint arXiv:2105.07346.

Yildiz, K. (2020). Identiﬁcation of wool and mohair ﬁbres

with texture feature extraction and deep learning. IET

Image Processing, 14(2):348–353.

Zhang, H., Li, A., Guo, J., and Guo, Y. (2020). Hybrid mod-

els for open set recognition. In Vedaldi, A., Bischof,

H., Brox, T., and Frahm, J.-M., editors, Computer Vi-

sion - ECCV 2020, pages 102–117, Cham. Springer

International Publishing.

Zhang, L. and Ainsworth, W. (2005). Microscope analysis

of animal ﬁber blends - training of operators. Techni-

cal report.

Zisselman, E. and Tamar, A. (2020). Deep residual ﬂow

for out of distribution detection. In Proceedings of

the IEEE/CVF Conference on Computer Vision and

Pattern Recognition (CVPR).

Zoccola, M., Lu, N., Mossotti, R., Innocenti, R., and Mon-

tarsolo, A. (2013). Identiﬁcation of wool, cashmere,

yak, and angora rabbit ﬁbers and quantitative deter-

mination of wool and cashmere in blend: a near

infrared spectroscopy study. Fibers and Polymers,

14(8):1283–1289.

Animal Fiber Identiﬁcation under the Open Set Condition