Professional Documents
Culture Documents
Textbook Advanced Concepts For Intelligent Vision Systems 19Th International Conference Acivs 2018 Poitiers France September 24 27 2018 Proceedings Jacques Blanc Talon Ebook All Chapter PDF
Textbook Advanced Concepts For Intelligent Vision Systems 19Th International Conference Acivs 2018 Poitiers France September 24 27 2018 Proceedings Jacques Blanc Talon Ebook All Chapter PDF
https://textbookfull.com/product/advanced-concepts-for-
intelligent-vision-systems-17th-international-conference-
acivs-2016-lecce-italy-october-24-27-2016-proceedings-1st-
edition-jacques-blanc-talon/
https://textbookfull.com/product/reachability-problems-12th-
international-conference-rp-2018-marseille-france-
september-24-26-2018-proceedings-igor-potapov/
Knowledge and Systems Sciences 19th International
Symposium KSS 2018 Tokyo Japan November 25 27 2018
Proceedings Jian Chen
https://textbookfull.com/product/knowledge-and-systems-
sciences-19th-international-symposium-kss-2018-tokyo-japan-
november-25-27-2018-proceedings-jian-chen/
https://textbookfull.com/product/computer-vision-and-graphics-
international-conference-iccvg-2018-warsaw-poland-
september-17-19-2018-proceedings-leszek-j-chmielewski/
https://textbookfull.com/product/database-systems-for-advanced-
applications-25th-international-conference-dasfaa-2020-jeju-
south-korea-september-24-27-2020-proceedings-part-i-yunmook-nah/
https://textbookfull.com/product/database-systems-for-advanced-
applications-25th-international-conference-dasfaa-2020-jeju-
south-korea-september-24-27-2020-proceedings-part-ii-yunmook-nah/
Jacques Blanc-Talon · David Helbert
Wilfried Philips · Dan Popescu
Paul Scheunders (Eds.)
Advanced Concepts
LNCS 11182
for Intelligent
Vision Systems
19th International Conference, ACIVS 2018
Poitiers, France, September 24–27, 2018
Proceedings
123
Lecture Notes in Computer Science 11182
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zurich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology Madras, Chennai, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7412
Jacques Blanc-Talon David Helbert
•
Advanced Concepts
for Intelligent
Vision Systems
19th International Conference, ACIVS 2018
Poitiers, France, September 24–27, 2018
Proceedings
123
Editors
Jacques Blanc-Talon Dan Popescu
DGA CSIRO-ICT Centre
Bagneux Canberra, ACT
France Australia
David Helbert Paul Scheunders
Laboratoire XLIM University of Antwerp
Futuroscope Chasseneuil Cedex Wilrijk
France Belgium
Wilfried Philips
Ghent University
Ghent
Belgium
LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
These proceedings gather the selected papers of the Advanced Concepts for Intelligent
Vision Systems (ACIVS) Conference, which was held in Poitiers, France, during
September 24–27, 2018.
This event was the 19th ACIVS. Since the first event in Germany in 1999, ACIVS
has become a larger and independent scientific conference. However, the seminal
distinctive governance rules have been maintained:
– To update the conference scope on a yearly basis. While keeping a technical
backbone (the classic low-level image processing techniques), we have introduced
topics of interest such as video analysis, segmentation, classification, remote
sensing, biometrics, deep learning, and image and video compression, restoration
and reconstruction. In addition, speakers usually give invited talks on hot issues.
– To remain a single-track conference in order to promote scientific exchanges among
the audience.
– To grant oral presentations a duration of 25 minutes and published papers a length
of 12 pages, which is significantly different from most other conferences.
The second and third items entail a complex management of the conference; in
particular, the number of time slots is rather small. Although the selection between the
two presentation formats is primarily determined by the need to compose a
well-balanced program, papers presented during plenary and poster sessions enjoy the
same importance and publication format.
The first item is strengthened by the notoriety of ACIVS, which has been growing
over the years: official Springer records show a cumulated number of downloads on
August 1, 2018, of more than 550,000 (for ACIVS 2005–2016 only).
The regular sessions also included a couple of invited talks by Yuliya Tarabalka
(Inria Sophia Antipolis, France) and Mihai Datcu (DLR, Germany). We would like to
thank all of them for enhancing the technical program with their presentations.
ACIVS attracted submissions from many different countries, mostly from Europe,
but also from the rest of the world: Belgium, China, Czech Republic, Finland, France,
Germany, India, Italy, Portugal, Korea, Romania, Saudi Arabia, Spain, the UK, the
USA, and Vietnam.
From 91 submissions, 36 were selected for oral presentation and 16 as posters. The
paper submission and review procedure was carried out electronically and a minimum
of two reviewers were assigned to each paper. A large and energetic Program Com-
mittee, helped by additional reviewers, as listed on the following pages, completed the
long and demanding reviewing process. We would like to thank all of them for their
timely and high-quality reviews, achieved in quite a short time and during the summer
holidays.
Finally, we would like to thank all the participants who trusted in our ability to
organize this conference for the 19th time. We hope they attended a different and
VI Preface
stimulating scientific event and that they enjoyed the atmosphere of the various ACIVS
social events in the city of Poitiers.
As mentioned, a conference like ACIVS would not be feasible without the concerted
effort of many people and the support of various institutions. We are indebted to the
local organizers Pascal Bourdon, David Helbert, Mohamed-Chaker Larabi, François
Lecellier, Benoit Tremblais, and Thierry Urruty, for having smoothed all the harsh
practical details of an event venue.
Steering Committee
Jacques Blanc-Talon DGA, France
David Helbert University of Poitiers, France
Wilfried Philips Ghent University - imec, Belgium
Dan Popescu CSIRO Data 61, Australia
Paul Scheunders University of Antwerp, Belgium
Organizing Committee
Pascal Bourdon XLIM, France
David Helbert University of Poitiers, France
Mohamed-Chaker Larabi XLIM, France
François Lecellier XLIM, France
Benoit Tremblais XLIM, France
Thierry Urruty XLIM, France
Program Committee
Hojjat Adeli Ohio State University, USA
Syed Afaq Shah The University of Western Australia, Australia
Hamid Aghajan Ghent University - imec, Belgium
Edoardo Ardizzone University of Palermo, Italy
Antonis Argyros University of Crete, Greece
George Bebis University of Nevada, USA
Fabio Bellavia University of Florence, Italy
Jenny Benois-Pineau University of Bordeaux, France
Dominique Béréziat Université Pierre et Marie Curie, France
Yannick Berthoumieu Bordeaux INP, France
Janus Bobulski Czestochowa University of Technology, Poland
Philippe Bolon University of Savoie, France
Egor Bondarev Technische Universiteit Eindhoven, The Netherlands
Don Bone University of Technology Sydney, Australia
Adrian Bors University of York, UK
Salah Bourennane Ecole Centrale de Marseille, France
Catarina Brites Instituto Superior Técnico, Portugal
Vittoria Bruni University of Rome La Sapienza, Italy
Dumitru Burdescu University of Craiova, Romania
VIII Organization
Additional Reviewers
Syed Afaq Shah The University of Western Australia, Australia
Hamid Aghajan Ghent University - imec, Belgium
Roxana Agrigoroaie U2IS, ENSTA ParisTech, France
Edoardo Ardizzone University of Palermo, Italy
Antonis Argyros University of Crete, Greece
George Bebis University of Nevada, USA
Fabio Bellavia University of Florence, Italy
X Organization
Video Analysis
Remote Sensing
Biometrics
Deep Learning
1 Introduction
Activity recognition, movement prediction (see Fig. 1), and detection of abnor-
mal behaviors are critical tasks in surveillance applications, requiring appropriate
models to describe them. Since many of the surveillance settings are based on
far-field cameras that do not provide detailed information about the pedestrians
in a video scene, several of these models are based on the characterization of
the trajectories performed by the pedestrians. In fact, trajectories allow us to
identify and collect statistics of typical motions, activities, and interactions in a
video scene [1]. Consequently, having a reliable description of the possible tra-
jectories is of major importance. However, trajectory modeling is a challenging
problem due to its great spatial and temporal variability.
c Springer Nature Switzerland AG 2018
J. Blanc-Talon et al. (Eds.): ACIVS 2018, LNCS 11182, pp. 3–13, 2018.
https://doi.org/10.1007/978-3-030-01449-0_1
4 C. Barata et al.
Fig. 1. Pedestrian trajectory: past (yellow) and possible future motions (green). (Color
figure online)
Our work is related to [9] where vector fields are √ used √ to model pedestrian
trajectories. This method assumes that a grid with n × n nodes is defined in
the image domain. Each node is characterized by the following parameters: (i) a
set of motion vector fields (velocity vectors), (ii) a field of isotropic covariance
matrices, and (iii) a switching probability matrix that governs the transition
between fields in that node. Only one motion field is active at each trajectory
point and the transition between them depends on the spatial location of the
pedestrian in the grid. This method has been show to efficiently characterize tra-
jectories in the task of activity recognition [9]. However, it has some limitations:
(i) the parameter estimation is performed using the EM algorithm that strongly
depends on the initialization, and may encourage wrong estimates; and (ii) the
fields were assumed to be dense and estimated in nodes where information was
not available. Recently, a new formulation was proposed to deal with the lat-
ter issue: imposing a sparsity constraint to the vector fields [10]. However, this
work failed to demonstrate the impact of such strategy at describing pedestrian
trajectories. Moreover, to the best of our knowledge, none of the vector fields
Improving a Switched Vector Field Model 5
models has ever been used to predict future positions in a trajectory. We believe
that this is a critical task in surveillance applications.
This paper provides valuable and comprehensive contributions, namely: (i)
provides an effective way on how to initialize the vector fields in the EM algo-
rithm, (ii) evaluates different vector fields formulations, and (iii) provides a
new methodology for trajectory prediction, using the space-varying dynamical
model. For the first item, we perform a systematic study comprising the follow-
ing three strategies: (a) random, (b) uniform, and (c) clustering initialization,
thus not jeopardizing deficient estimates related to poor initializations. For the
second item, we compare the vector fields formulations of [9,10]. The proposed
prediction strategy is used to evaluate the initialization and type of fields.
n
B(x) = B i φi (x), (4)
i=1
where Tk is the dictionary of node velocities Tki associated with the k − th field,
B i is the switching matrix associated with node gi , Φ(x) is a N × 1 normalized
sparse vector of known interpolation coefficients, and φi (x) is its i − th element.
6 C. Barata et al.
where θ = (T, B, σ 2 ) denotes the complete set of model parameters. From (5),
it is clear that the sequence of active fields k = (k1 , ..., kL ) is modeled as a
realization of a first order Markov process, with some initial (known) distribution
P (k1 ) and a space-varying transition matrix, i.e., Bkt−1 ,kt (xt−1 ). In the following
section we address the estimation of the model parameters.
K
where log p(θ) = k=1 log p(θk ) is the log prior distribution. Assuming indepen-
dence between the various parameters, it is possible to further decompose the
prior as follows
K
K
log p(θ) = log p(Tk ) + log p(σk2 ) + log p(B). (7)
k=1 k=1
The complete likelihood log p(X, K|θ) may also be easily defined from (5),
by computing a sum over all the trajectories. However, its marginalizing over all
the admissible labels K cannot be computed.
The aforementioned limitation may be addressed using the Expectation-
Maximization (EM) method to estimate the model parameters. This amounts
to iteratively maximizing the auxiliary function
with respect to θ and taking into account the available estimates θ from the
previous iteration.
Improving a Switched Vector Field Model 7
Using (5) and replacing into (8) we obtain the following expression (disre-
garding the constants)
L
S K
U (θ, θ ) = −
(s)
wk (t) log N (xt |xt−1 + Tkt (xt−1 ), σk2t I)
s=1 t=2 k=1
S
L
K
+ wp,q
(s)
(t) log Bkt−1 ,kt (xt−1 )
s=1 t=2 p,q=1
K
K
+ log p(Tk ) + log p(σk ) + log p(B), (9)
k=1 k=1
These weights are computed in the E-step, using the forward-backward algorithm
proposed in [12]. The optimization with respect to θ is performed in the M-step,
and involves a separate maximization with respect to T, B, and σ 2 [9].
As discussed in [10] the prior log p(θ) may be used to incorporate a priori
knowledge about the parameters, as well as to enforce specific properties. In
particular, efforts have been made to define a prior distribution p(T) that could
lead to smooth and sparse fields:
4 Experimental Setup
The goal of this work is to provide an extensive experimental study on the role of
both initialization and the prior p(T) in the estimation of the motion models. To
conduct such a study it is necessary to (i) identify the types of initialization and
priors to be tested; (ii) select one or more datasets to conduct the experiments;
8 C. Barata et al.
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 2. Experimental datasets: synthetic examples (1st row) and real pedestrian tra-
jectories (2nd row).
and (iii) define an evaluation metric. In this section we address each of these
requirements.
(i) Initialization: The EM-algorithm, which is used in the estimation of
the parameters, strongly depends on initialization. Thus, we will evaluate three
types of initialization: (i) random, where each node of a field is initialized with
a random velocity vector; (ii) uniform, where each node of a field is initialized
with the same (random) velocity vector; and (iii) clustered, where we perform a
preliminary clustering of the velocity values and set the number of centroids to
be equal to the number of fields K.
(ii) Prior: Another goal of this paper is to provide a comprehensive study
of the role of the fields’ prior (10). This means that for each of the aforemen-
tioned initializations, we train several models using different combinations of
α ∈ {0.5, 1, 2}, β ∈ {0, 0.05, 0.1, 0.15}, and p ∈ {1, 2}, leading to a total of 21
models per initialization type.
(iii) Datasets: Our study is conducted using four datasets (two synthetic
and two real). The first synthetic case comprises K = 3 motion types and
100 randomly generated trajectories, as is exemplified in Fig. 2 (1st row-left).
The trajectories were generated according to the following procedure. First, the
starting point was randomly chosen between a small region around [0, 0.9]T and
[0.5, 0]T . Conditioned on the selected point, the motion regime can either be a
down motion with a slope of 45◦ or an up motion. When the trajectories reach
Improving a Switched Vector Field Model 9
the center of the image there is a probability of switching between these motions
and a third one, oriented form left to right.
1
L
Table 1. Dataset description. L̄ is the average trajectory length and v̄ = L t=1 xt −
xt−1 22 is the average displacement.
goal, we apply the forward-backward algorithm [12] used in the E-step of model
estimation, namely the forward part, that works as follows. First, lets consider
the past, i.e., t ≤ t0 , and assume the existence of a set of forward variables
such that α(t) = [α1 (t), ..., αK (t)] is a vector that stores all of the forward
probabilities of time instant t. For a given past instant, α(t) is estimated as
follows:
For t > t0 the real position x(t) is unknown. Therefore, matrix D can not be
computed and the estimation of the forward probabilities α̂(t) is given by:
where
x̂t = x̂t−1 + Tkt∗ (x̂t−1 ) + 0. (15)
kt∗ denotes the active field at each instant t > t0 , and is defined as: k ∗ =
maxk αk (t).
We use the displacement error (DE) as the evaluation metric:
L−δ
DE = x(t0 + δ) − x̂(t0 + δ)22 , (16)
t0 =1
5 Results
In Fig. 3 we show the best performances for each dataset (synthetics 1 and
2, Stairway, and Campus II). These results were obtained using leave-one-
trajectory-out cross validation and the performances were computed for each
pair (initialization,prior ). Here dense means that β = 0, while l2 and l1 corre-
spond to p = 2, 1, in (10). Table 2 shows the best overall performances, as well
as the combination of model parameters and initialization that led to it. These
results demonstrate that the proposed prediction strategy is able to successfully
the estimate the position of a pedestrian δ = 5 time steps ahead, achieving a
low DE in all datasets.
Improving a Switched Vector Field Model 11
0.014
Syn. 1 + dense
0.012 Syn.1 + l2
Syn. 1 + l1
0.01 Syn. 2 + dense
Syn. 2 + l2
0.008 Syn. 2 + l1
DE
Stairs. + dense
0.006 Stairs. + l2
Stairs + l1
0.004 Campus + dense
Campus + l2
0.002 Campus + l1
0
Random Uniform Clustering
Initialization
Fig. 3. Experimental results for synthetic examples (red and blue) and real pedestrian
trajectories (green and magenta). (Color figure online)
Table 2. Best experimental results out of 63 possible configurations for each dataset.
A more thorough analysis of the results suggest that the initialization method
has a significant impact in the synthetic and Stairway datasets, while the per-
formance in the Campus II remains almost unchangeable. The role of the ini-
tialization is particularly notorious for synthetic example 2 (the roundabout),
where velocity clustering leads to a large increase in the DE. This higher error
explained by the existence of a circular movement that is inefficiently modeled
by velocity clustering.
Enforcing sparsity in the estimation of the fields (p = 1) leads to improved
prediction results for the synthetic data. However, the opposite is observed for
the real datasets, where dense fields achieve the best results. Such disparity may
be due to the proportionally higher variability that exists in pedestrian velocities,
when compared with the synthetic datasets (see Table 1). Setting p = 2 in (10)
12 C. Barata et al.
has some impact in the performance, but it is clearly outperformed by the other
priors. Regarding the prior regularization constants, α and β, it seems that
setting α = 0.5 and β = 0.05 is the preferred combination for the sparse prior
in three out of the four datasets. In the case of the dense prior, we observe that
setting α to a higher value leads to the best results for the real datasets, while
no trend is observed for the synthetic ones.
6 Conclusions
Efficiently modeling the trajectories of pedestrians is critical for most surveil-
lance setups. However, several of the models proposed to tackle this issue rely
on complex methods that require the tunning of multiple parameters. In this
paper we addressed this issue, taking as starting point a motion model based on
switching motion fields.
Our contributions are two-folded. First we proposed a trajectory prediction
strategy for the aforementioned method, which allows a qualitative evaluation of
the method. Second, we conducted a thorough study of the model parameters,
namely initialization, velocity priors, and constants, in order to obtain insights
of their role. Both contributions are original of this paper. Experiments were
conducted in four datasets and the obtained results are relevant and promising.
Future work should rely on exploring the applicability of the motion fields to
other tasks, such as the detection of abnormal trajectories or activities.
Acknowledgments. This work was supported by the FCT project and plurianual
funding: [PTDC/EE- IPRO/0426/2014], [UID/EEA/50009/2013].
References
1. Aggarwal, J., Ryoo, M.: Human activity analysis: a review. ACM Comput. Surv.
43(3), 16 (2011)
2. Kim, K., Lee, D., Essa, I.: Gaussian process regression flow for analysis of motion
trajectories. In: Proceedings of IEEE International Conference Computer Vision
(ICCV) (2011)
3. Ellis, D., Sommerlade, E., Reid, I.: Modelling pedestrian trajectory patterns with
Gaussian processes. In: Proceedings IEEE International Conference Computer
Vision (ICCV) Workshops (2009)
4. Bastani, V., Marcenaro, L., Regazzoni, C.S.: Online nonparametric bayesian activ-
ity mining and analysis from surveillance video. IEEE Trans. Image Process. 25(5),
2089–2102 (2016)
5. Ferreira, N., Klosowski, J.T., Scheidegger, C.E., Silva, C.T.: Vector field k-means:
clustering trajectories by fitting multiple vector fields. In: Eurographics Conference
on Visualization, vol. 32(3), pp. 201–210 (2013)
6. Bastani, V., Marcenaro, L., Regazzoni, C.: Unsupervised trajectory pattern clas-
sification using hierarchical Dirichlet process mixture hidden Markov model. In:
Proceedings of IEEE International Workshop Machine Learning for Signal Pro-
cessing (MLSP), p. 16 (2014)
Improving a Switched Vector Field Model 13
7. Hu, W., Li, X., Tian, G., Maybank, S., Zhang, Z.: An incremental DPMM-based
method for trajectory clustering, modeling, and retrieval. IEEE Trans. Pattern
Anal. Mach. Intell. 35(5), 10511065 (2013)
8. Lin, D.: Online learning of nonparametric mixture models via sequential variational
approximation. In: Advances in Neural Information Processing Systems, pp. 395–
403 (2013)
9. Nascimento, J.C., Figueiredo, M.A.T., Marques, J.S.: Activity recognition using
mixture of vector fields. IEEE Trans. Image Process. 22(5), 1712–1725 (2013)
10. Barata, C., Nascimento, J.C., Marques, J.S.: A sparse approach to pedestrian
trajectory modeling using multiple motion fields. In: International Conference on
Image Processing (2017)
11. Figueiredo, M.A.T., Jain, A.K.: Unsupervised learning of finite mixture models.
IEEE Trans. Pattern Anal. Mach. Intell. 24(3), 381–396 (2002)
12. Rabiner, L.R.: A tutorial on Hidden Markov Models and selected applications in
speech recognition. Proc. IEEE 77(2), 257–286 (1989)
13. Alahi, A., et al.: Learning to predict human behaviour in crowded scenes. In: Group
and Crowd Behavior for Computer Vision, pp. 183–207 (2017)
Another random document with
no related content on Scribd:
Amblyoponides, 132, 180
Ambrosia, 295
Ambryllis, 409
Amicta quadrangularis, 394
Ammophila, 111;
A. affinis, 111;
A. hirsuta, 111;
A. holosericea, 111
Amphicyrtides, 242
Amphidasis betularia, 412, 414
Amphiodont, 193
Amphipneustic, 450
Amphizoa lecontei, 207
Amphizoidae, 207
Ampulex compressa, 114, 115;
A. ruficornis, 115;
A. sibirica, 114
Ampulicides, 114 f., 169
Amycterides, 291
Anal armature, 328, 416
Anal nervures, 318
Anaphe, 376
Anaspini, 267
Anaspis, 268
Anatomy—see External Structure and Internal Anatomy
Anchor-process, 459
Ancient, Lepidopteron, 435—see also Primitive
Ancylolominae, 425
Andrena, 23, 25, 30, 301, 303;
hair of, 11;
A. labialis, 488;
A. nigroaenea, 23;
A. ovina, 30
Andrenidae, 20
Andrenides, 23
Andrenimorpha, 388
Andrenoides, 20
Androconia, 331 f.
Anergates atratulus, 160 f.
Angelitos, 63
Anisopteryx aescularia, 411
Anisotoma cinnamomea, 222
Anisotomidae, 223
Anisotomides, 222
Anlagen, 143
Anobiides, 247
Anobium, 254;
A. paniceum, 247;
A. striatum, 248;
A. tessellatum, 248
Anochetus ghiliani, 174
Anomma, 178
Anophthalmus, 205
Anoplura, 599 f.
Anosia erippus, A. menippe, 345;
A. plexippus, 307;
larva, 324;
pupa, 327
Antennae;
of Belostoma, 566;
of flies, 441;
of Lamellicornia, 191;
of Lepidoptera, 307;
of butterflies, 340, 341;
of Sphingidae, 380
Antherophagus, 235
Anthicidae, 266
Anthidium, 45;
A. bellicosum, 47;
A. diadema, 45;
A. manicatum, 45;
A. septemdentatum, 47;
A. strigatum, 30
Anthocopa, 51
Anthocoridae, 560
Anthomyia angustifrons, A. brassicae, A. cana, 506
Anthomyiidae, 506
Anthophila, 10 f.
Anthophora, 32, 33;
destroyer of, 272, 274;
A. personata, 33;
A. pilipes, 33;
proboscis of, 17
Anthothrips aculeata, 530
Anthracides, 486
Anthrax, 486 f.;
A. fenestralis, 489;
A. trifasciata, 44
Anthrenus fasciatus, 241
Anthribidae, 278, 290
Antisquama, 448
Antitegula, 447
Ant-plant, 138, 139, 168
Ants, 131 f.;
and Aphidae, 590;
and caterpillars, 356
Ants'-nest, Insects, 200, 213, 221, 223, 224, 225, 231, 236, 240,
548;
larva, 501, 502
Anus, 314, 320
Aorta, 320
Aortal chamber, 320
Apate capucina, 246
Apatela, 418
Apathus = Psithyrus, q.v.
Apatidae, 246
Apatura;
larvae, 354;
A. iris, 344
Apaturides, 352
Aphaenogaster, 164, 165, 221;
A. arenarius, 164;
A. barbarus, 131, 164;
A. structor 164, 240
Aphaniptera, 522 f.
Aphanocephalus, 228
Aphelocheirus, 565
Aphidae, 581 f.;
and ants, 181
Aphidiides, 590
Aphis maidi-radicis, 584
Aphomia sociella, 424
Apidae, 10 f., 20, 32
Apioceridae, 492
Apis, 53;
A. adansonii, 69;
A. domestica, 68;
A. dorsata, 69;
A. fasciata, A. ligustica, 68;
A. mellifica, 65 f.;
feet of queens and workers, 69;
ligula of, 16;
worker and hairs, 12
Apoda, 402;
A. testudo, etc., 401, 402
Apodidae, 402
Apoica pallida, nest, 83
Aporia crataegi, 322
Appetite, 491
Apterogyna, 96
Apterona, 393, 394;
A. crenulella, var. helix, 395
Apterous, beetles, 187, 263;
females, 95, 96, 140, 171, 174, 315, 392, 393, 407, 413, 430,
592—see also Workers;
Insects, 95, 96, 474, 495, 496, 505, 518, 531, 581;
males, 140, 160, 161, 172, 585, 594
Aquatic;
caterpillar, 377;
cocoons, 280;
larva, 421 f., 425, 504;
pupa, 423
Aradidae, 550
Aradus orientalis, 550
Araeocerus, 290
Araschnia levana, A. prorsa, 353
Arbelidae, 369, 396
Arceina, 373
Archiapidae, 21, 22
Archiapides, 21 f.
Arctia caja, 308;
A. villica, 410
Arctiidae, 370, 404, 408, 410
Argiva, 414
Argynnis, larvae, 354;
A. paphia, androconia, 332
Argyromoeba sinuata, 76;
A. trifasciata, 486
Arista, 442
Army-worm, 416
Aromia moschata, 188
Arthropterus, 214
Asclera caerulea, 267
Ascodipteron, 520
Asemorhoptrum lippulum, 160
Asilidae, 491
Asilus, 492;
A. crabroniformis, 441
Asparagus-beetle, 281
Aspidiotus camelliae, 592;
A. nerii, 595, development of, 596
Aspidiphorus, 246
Aspidomorpha, pupa, 283, 284
Association, of Anergates and Tetramorium, 160;
of ants and other Insects etc., 180 f.;
of Formica and Formicoxenus, 159;
of Strongylognathus and Tetramorium, 162;
of Tomognathus and Leptothorax, 161
Astata boops, 119
Astatides, 119
Asteidae, 504
Astomella lindeni, 490
Astynomus, 285
Atemeles, 225;
and ant, 182
Athericerous, 441
Atherix, 481;
A. ibis, 480
Athous rhombeus, 257
Atractocerus, 254
Atrophy of mouth and stomach, 310
Atta, 137, 164 n., 165, 502
Attacus, 373;
A. atlas, 373
Attelabides, 291
Attini, 158, 159, 165
Attitude, 381, 384, 385, 388, 412, 413, 425, 429
Atylotus fulvus, 483
Autocrates aenea, 275
Automeris, 373
Axin, 598
Azteca, 158
Azygos oviduct, 321
Dacnides, 237
Dactylopius citri, 595
Daddy-long-legs, 471
Dakruma coccidivora, 424
Dalla Torre, Catalogue of Hymenoptera, 21
Danaides, 344, 347
Danaioid Heliconiidae, 346
Danais archippus, or plexippus, 345
Dances, 351, 464, 493, 554
Darwin, C., on Pelobius, 208
Darwin, F., on proboscis of Lepidoptera, 311 n.
Dascillidae, 243, 255
Dascillus cervinus, 255
Dasychira pudibunda, 408;
D. rossii, 407
Dasygastres, 20, 35 f.
Dasypoda hirtipes, 27