=Paper=
{{Paper
|id=Vol-3180/paper-178
|storemode=property
|title=1st Place Solution for FungiCLEF 2022 Competition: Fine-grained Open-set Fungi Recognition
|pdfUrl=https://ceur-ws.org/Vol-3180/paper-178.pdf
|volume=Vol-3180
|authors=Zihua Xiong,Yumeng Ruan,Yifei Hu,Yue Zhang,Yuke Zhu,Sheng Guo,Bing Han
|dblpUrl=https://dblp.org/rec/conf/clef/XiongRHZZ0H22
}}
==1st Place Solution for FungiCLEF 2022 Competition: Fine-grained Open-set Fungi Recognition==
https://ceur-ws.org/Vol-3180/paper-178.pdf
1st Place Solution for FungiCLEF 2022 Competition:
Fine-grained Open-set Fungi Recognition
Zihua Xiong, Yumeng Ruan, Yifei Hu, Yue Zhang, Yuke Zhu, Sheng Guo and
Bing Han
MYbank, Ant Group, China
Abstract
In this paper, we describe our method for Fine-Grained Fungi Recognition at FungiCLEF 2022, which
aims to recognize the fungi belonging to 1,604 known species and many other unknown species, termed as
a fine-grained, open-set machine learning problem. For the purpose of building a strong close-set classifier,
we taken MetaFormer [1] and ConvNext [2] as our strong baseline, then we applied hyper-parameter
tuning and some modern training techniques to improve it. To deal with long tailed class distribution
problem, we adapt the Seesaw Loss [3] to balance the training process between head classes and tail
classes. Furthermore, to avoid tail categories being misclassified as open-set categories, we intuitively
design a post process to alleviate the confusion. As a common practice, test time augmentations and model
ensemble are used. With all these techniques together, our method achieves superior mean π 1 score on test
set, that is 83.78% on public leaderboard, and 80.43% on private leaderboard which is the 1st place among
the participators. The code will be made available at https://github.com/guoshengcv/fgvc9_fungiclef.
Keywords
FungiCLEF, fungi recognition, long tail, open-set, fine-grained, classification
1. Introduction
FungiCLEF 2022 [4] is a competition held jointly by CLEF 2022 conference [5, 6] and FGVC9
workshop at CVPR 2022 conference. The competition release the train data based on Danish
Fungi 2020 [7], which aims at fine-grained fungi recognition. It includes both image and
meta-information such as habitat, substrate, time, longitude, latitude etc, and contains 295,938
samples belonging to 1,604 species. The competition also release test data which contains 59,420
observations with 118,676 images and 3,134 species, it includes meta-information as train data
but miss some attributes such as longitude and latitude. After data analyze, we found the category
distribution of the train dataset is long-tailed, as a result, in this work we tackle the competition
as a fine-grained, long-tailed, open-set classification task.
Different from common classification tasks that try to distinguish objects with large inter-class
variations, Fine-Grained Visual Classification (FGVC) aims at capturing the subtle difference
within similar categories, such as differentiating bird species, car types, etc. It is acknowledged
that FGVC is a challenging task due to small inter-class variations and large intra-class variations.
CLEF 2022: Conference and Labs of the Evaluation Forum, September 5β8, 2022, Bologna, Italy
$ xiongzihua.xzh@alibaba-inc.com (Z. Xiong); ruanyumeng.rym@alibaba-inc.com (Y. Ruan);
huyifei.hyf@alibaba-inc.com (Y. Hu); mosay.zy@alibaba-inc.com (Y. Zhang); felix.yk@alibaba-inc.com (Y. Zhu);
guosheng.guosheng@mybank.cn (S. Guo); hanbing.hanbing@antgroup.com (B. Han)
Β© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
Numerous methods for FGVC are mainly focused on modeling discriminative regions, such
as part-based model [8, 9, 10] and attention-based model [11, 12]. Recently, inspired by the
fact that human experts use meta-information to distinguish visually similar species, there are
many works [13, 14, 1, 15] utilize additional information to enhance fine-grained classification
performance. Among them, Metaformer [1] is the state-of-the-art work proposed recently, it
is a hybrid framework that convolution and transformer are both used. In this work, we taken
Metaformer as a strong baseline and improve it progressively.
Recently, transformers have leading the research in the filed of computer vision, starting from
Vision Transformer [16], there are many various transformer backbones achieve SoTA perfor-
mance in a wild range of vision tasks, such as Swin Transformer [17], CSwin Transformer [18],
etc. On the other side, ConvNext [2] is a pure convolution backbone, it applies modern training
techniques, macro and micro design of the network architecture, achieves comparable results
with transformers. In this work, in order to obtain models with distinct difference and enhance
the performance of model ensemble, we taken ConvNext as another baseline backbone.
In real world scenarios, the distribution of the categories is often long-tailed. It is well known
that major class will dominate the training process and suppress the performance of tail class.
Many works designed loss function to deal with the problem of long-tail classification, such
as Adaptive Class Suppression Loss [19], Equalization loss [20], Seesaw Loss [3], etc. In our
method, we utilize Seesaw Loss to dynamically balance the training process between head classes
and tail classes.
For practical application, it usually faces with open-set recognition challenge, the classifier
should not only recognize the classes which have been seen during training, but also notice that if
a instance comes from unknown classes. Motivated by [21], in this work, we trained the model on
known classes to obtain a good close-set classifier, and determine whether a instance belonging
to open-set based on the maximum value of itβs logit score vector. Furthermore, to avoid tail
categories being misclassified as open-set categories, we intuitively design a post process to
alleviate the confusion.
Our main contributions in FungiCLEF 2022 competition can be summarized as follows:
β’ We take Metaformer and ConvNext as our strong baseline, then we apply hyper-parameter
tuning and some modern training techniques to improve itβs performance on fungi dataset.
β’ We find category distribution of the fungi dataset is long-tailed, thus we adapt the Seesaw
Loss to balance the training process between head classes and tail classes, which lift up the
baseline model performance.
β’ To avoid tail categories being misclassified as open-set categories, we intuitively design a
post process to alleviate the confusion.
β’ Detailed ablation experiments have been done. With the techniques above, we achieve
superior performance.
2. Approach
Motivated by [21], we divide the fine-grained, open-set recognition problem into two parts. Firstly
we are attended to lift up the close-set recognition performance, including network architecture,
Figure 1: The overall of our apporach. We trained MetaFormer and Convnext with various
settings, during testing process, model snsemble and post process are used.
hyper-parameter tuning and long tail loss. Then we analyze the data and logit score frequency
distributions, and design the post process for open-set recognition. Model ensemble is done by
averaging the output logits from multiple models.
2.1. Overview of the Approach
As shown in Figure 1, we taken MetaFormer [1] and ConvNext [2] as our initial baseline.
MetaFormer is a hybrid framework that combines convolution and vision transformer, it also
proposes a simple and effective solution for adding meta-information using the transformer layer.
In our approach, we directly use MetaFormer and modify the input of meta-information. We
perform the mapping [ππππ‘β, πππ¦] β [π ππ( 2πππππ‘β 12 ), πππ ( 2πππππ‘β
12 ), π ππ( 2ππππ¦ 2ππππ¦
31 ), πππ ( 31 )]
to encode temporal information. We use one-hot encoding to encode category meta-information
such as countryCode, Substrate and Habitat. To enhance the model diversity for later model
ensemble, we use ConvNext as another network architecture. We apply hyper-parameter tuning
to improve their performance, and we will illustrate the ablation studies in Sec 3.2 to show the
progressive process.
2.2. Loss for Long Tail Classification
It is known that instances from head categories dominate the training process, the biased learning
lead to misclassification for tail categories. In this work, we borrow the idea from Seesaw Loss [3]
to alleviate this problem. During training process, Seesaw Loss dynamically balances positive
and negative gradients for each category with a dynamic factor, it reformulate the Cross Entropy
loss as
πΆ
βοΈ
πΏπ πππ ππ€ (π§) = β π¦π log(ΜοΈ
ππ ),
π=1 (1)
π π§π
with πΜοΈπ = βοΈπΆ .
π§π + ππ§π
πΜΈ=π πππ π
where π¦ is the category label, usually represented by one-hot, π§ is the outputs of model, πΜοΈ is
the probability calculated by πππ π‘πππ₯(π§) with a dynamic factor π. For more detail, please refer
to Seesaw Loss [3].
2.3. Post Process
The post process is intuitively designed based on several observations, and applied on final
ensemble results. In order to be as clear as possible the process, we write the post process in
python-style pseudocode, refer to the Algorithm 1. We detailed it in the following.
Threshold for selecting open-set samples. It is acknowledged that the predict confidence
score of open-set samples are relatively low. This phenomenon can be used as the criterion
for their recognition. Specifically, we draw the logit (the direct output of the model) frequency
distribution of both validation set and test set, shown in Figure 2. As the open set samples are
only contained in the test set, we can compare the low confidence areas of the two distributions to
approximately get the logit threshold for open set samples. For example, we can set threshold
to 5 as an approximate for the Figure 2. On this basis, we have draw a rough conclusion that
the test set contains approximate 1000 βΌ 2000 open set samples. It should be noted that this
rough conclusion may be wrong, since we have no information about the reality that how many
open-set samples in test set. Despite of it, in the rest of the experiments, we set the samples with
top-k lowest confidence as the open set, the value of π is set based on this rough conclusion. As
shown from line 7 to line 9 in Algorithm 1, we adjust the threshold to obtain open-set samples.
For different experiment setting and model, it is hard and needless to have exactly same number
of open-set samples, experimentally, we set k to βΌ 1000 at first, and adjust it to βΌ 1500 at the
final based on the public test set performance.
Alleviate the influence of microscopy images. In the test set, we find that one test sample may
contains several images as shown in Figure 3. For such case, we average the model outputs
of them to get the confidence and use argmax to get predicted category. During the above
process, we also find that there are images showing huge visual discrepancy with the majority.
Specifically, we find that some test samples contain microscopy images such as sample_c shown
in Figure 3, these microscopy images tend to produce low confidence due to little training data, it
will influence the naive average strategy. To cope with this problem, we delicately design the post
process. As shown in Algorithm 1, from line 29 to line 35, if the maximum logit of averaged
outputs is lower than a certain threshold, we will look into the maximum logit of all images from
a test sample, if it is greater than a certain threshold, we will get the corresponding category as
the prediction. In this case, we think that the test samples with low averaged outputs may be
caused by containing too many microscopy images, the high confidence prediction of one image
from test sample is sufficient to infer the category, and the test sample should not be considered
as open-set categories.
Figure 2: Logit score frequency distribution of val and test. We draw the Logit score frequency
distribution on both validation and test set, with the output logits of a single model.
Figure 3: Selected test samples. We found many test samples contain several images, we
predict the category of the test sample by averaging the model outputs of them. We also notice
that some test samples such as sample_c contains one image pictured from natural environment,
and the other images come from microscope view, it will disturb the average results.
Distinguish tail categories and open-set categories. We put the tail categories that are never
been predicted by the model into hard tail categories, we argue that there are many hard tail
categories misclassified as open-set categories. To deal with the problem above, we design the
post process, refer to the Algorithm 1. From line 18 to line 27, to avoid the misclassification, we
mining hard tail categories from top-3 predictions with low threshold filtering.
Algorithm 1 Pseudocode of Post Process in a python-like style.
1 # logit_scores: logit scores of a test sample
2 # high_t: high threshold
3 # low_t: low threshold
4 # hard_tail_categories: list of hard tail categories
5
6 # we dynamically adjust the threshold to have certain number of open-set samples
7 high_t = 9.8
8 low_t = high_t - 0.7
9
10 # we average logit scores of all images from test sample, then get top 3 classes and scores
11 mean_scores = mean(logit_scores, dim=0)
12 top1_cls, top2_cls, top3_cls, top1_score, top2_score, top3_score = topk(mean_scores, topk=3)
13
14 # we maximum logit scores of all images from test sample
15 max_scores = max(logit_scores, dim=0)
16 max_cls, max_score = argmax(max_scores)
17
18 # we distinguish tail categories and open-set categories
19 # by mining hard tail categories from top-3 predictions
20 if top1_score > low_t and top1_cls in hard_tail_categories:
21 predicted_cls = top1_cls
22 elif max_score > low_t and max_cls in hard_tail_categories:
23 predicted_cls = max_cls
24 elif top2_score > low_t and top2_cls in hard_tail_categories and top1_score - top2_score < 1.2:
25 predicted_cls = top2_cls
26 elif top3_score > low_t and top3_cls in hard_tail_categories and top1_score - top3_score < 1.2:
27 predicted_cls = top3_cls
28
29 # to alleviate the influence of microscopy images
30 if top1_score > high_t:
31 predicted_cls = top1_cls
32 # if the maximum logit of averaged outputs is lower than a certain threshold,
33 # we will look into the maximum logit of all images from a test sample
34 elif max_score > 15:
35 predicted_cls = max_cls
36 # recognized as open-set categories
37 else:
38 predicted_cls = -1
3. Experiments
In this section, we first elaborate on the implementation and training details. Then we introduce
ablation studies on loss functions and bag of training settings. Then we list some other attempts
and itβs results. Finally we study on different test time augmentations, and show the effectiveness
of post process for tail categories recognition and open-set categories recognition.
3.1. Implementation Details
We trained the model on Danish Fungi 2020 dataset [7] which contains 295,938 training images
belonging to 1,604 species observed mostly in Denmark, the dataset has been divided into train
and validation set. We use both train and validation for training in most settings. We report the
results on test set which contains 59,420 observations with 118,676 images and 3,134 species.
The test set is divided into 2 parts, the public set contains 20% of the data, the private set contains
80% of the data. As the performance of open-set recognition affects the mean π 1 score, to make
relative fair comparison in ablation studies, based on the observation illustrated in Sec 2.3, we
utilize threshold to select βΌ1000 samples which have low confidence score as open-set samples
for most experiments. We conduct all the experiments with Tesla V100 (32G). We use AdamW
optimizer with cosine learning scheduler, initialize the learning rate to 5πβ5 and scale it by batch
size, we follow most of the augmentation and regularization strategies of [17] in training.
Table 1
MetaFormer-0 baseline.
loss batch size accumulate steps epochs mixup train+val public mean π 1 private mean π 1
Soft Target CE 32 1 100 yes no 78.76% 74.26%
Table 2
Mean π 1 score on public/private test set with different losses and MetaFormer-0 as backbone.
loss batch size accumulate steps epochs mixup train+val public mean π 1 private mean π 1
Soft Target CE 32 1 32 yes no 71.49% 67.6%
Label Smoothing CE 32 1 32 no no 76.9% 72.48%
Label Smoothing CE 64 3 32 no yes 79.45% 75.67%
Seesaw Loss 64 3 32 no yes 79.79% 76.15%
Table 3
Mean π 1 score on public/private test set with different batch size and MetaFormer-0 as backbone.
loss batch size accumulate steps epochs mixup train+val public mean π 1 private mean π 1
Label Smoothing CE 32 1 32 no no 76.90% 72.48%
Label Smoothing CE 64 1 32 no no 77.49% 74.27%
Table 4
Mean π 1 score on public/private test set with different accumulate steps.
loss batch size accumulate steps epochs backbone train+val public mean π 1 private mean π 1
Seesaw Loss 64 3 32 MetaFormer-0 yes 79.79% 76.15%
Seesaw Loss 64 6 32 MetaFormer-0 yes 80.22% 76.90%
Seesaw Loss 32 3 64 MetaFormer-1 yes 81.67% 77.62%
Seesaw Loss 32 6 64 MetaFormer-1 yes 81.66% 77.94%
3.2. Ablation Studies
As shown in Table 1, we train MetaFormer-0 for 100 epochs, with Soft Target Cross Entropy loss
and mixup [22] augmentation to build our baseline. For ablation studies, it should be noted that
except the parameter to be compared, there are little other not consistent parameters, such as the
accumulate steps in last row in Table 5, we argue that it will not affects the conclusion largely.
Losses. As shown in Table 2, we compare different losses with several common augmentation
techniques. Specifically, we compare cross entropy loss with either mixup or label smoothing [23]
and Seesaw loss. They are all devoted to alleviate the long-tail problem in training. It is found that
label smooth converges faster than mixup in our experiments. The best performance is achieved
when Seesaw loss is adopted.
Batch size. Table 3 illustrates that larger batch size improves the performance. in detail, by
increasing batch size from 32 to 64, we improved π 1 score from 76.90% to 77.49% on public set,
consistently improve π 1 score from 72.48% to 74.27% on private set. Similar techniques is to
increase the accumulate steps. As shown in Table 4, enlarging accumulate steps improves mean
π 1 score in private test set consistently with MetaFormer-0 and MetaFormer-1.
Training epochs. We found the longer training epochs will not definitely improve the perfor-
mance. As shown in Table 5, for MetaFormer-0 and MetaFormer-2, it is consistent that proper
epochs is essential for better result. Following this line, we did not train models with dozens of
epochs such as 100 epochs and above.
Image size. Usually, training with larger image size improves the overall performance, especially
for fine-grained tasks. We use the 384 as the baseline image size and try several other larger
settings. As shown in Table 6, the larger image size 448 does not consistently bring improvements
Table 5
Mean π 1 score on public/private test set with different training epochs.
loss batch size accumulate steps epochs backbone train+val public mean π 1 private mean π 1
Seesaw Loss 64 3 32 MetaFormer-0 yes 79.79% 76.15%
Seesaw Loss 64 3 64 MetaFormer-0 yes 80.46% 77.01%
Seesaw Loss 64 3 100 MetaFormer-0 yes 80.18% 76.77%
Seesaw Loss 24 4 32 MetaFormer-2 yes 81.18% 77.56%
Seesaw Loss 24 4 48 MetaFormer-2 yes 82.04% 77.92%
Seesaw Loss 24 6 64 MetaFormer-2 yes 80.45% 77.63%
Table 6
Mean π 1 score on public/private test set with different image size and MetaFormer-1 as back-
bone.
image size batch size accumulate steps epoch public mean π 1 private mean π 1
384 32 6 64 81.76% 78.25%
448 20 6 64 80.79% 78.48%
Table 7
Mean π 1 score on public/private test set with different pretrain dataset and MetaFormer-2 as
backbone.
pretrain dataset batch size accumulate steps epochs public mean π 1 private mean π 1
herbarium 12 6 32 80.90% 77.37%
imagenet22k 12 8 48 81.47% 77.86%
inaturalist21 24 6 48 82.04% 77.92%
Table 8
Mean π 1 score on public/private test set with ConvNext-tiny, ConvNext-base and ConvNext-large
as backbone. Notice that we only use image data to train ConvNext.
batch size accumulate steps epochs +pseudo label backbone public mean π 1 private mean π 1
96 3 64 no convnext-tiny 76.93% 73.46%
32 4 64 no convnext-base 78.97% 75.46%
10 4 64 no convnext-large 79.15% 75.59%
24 6 80 yes convnext-large 80.65% 76.61%
on public test set. We blame it to the coupled training schedules with the image size, which we
do not investigate into it. Finally, we adopt the image size 384 in all settings. Nevertheless, the
performance with this baseline is satisfactory enough.
Pretrain dataset. We transfer MetaFormer-2 pretrained on different dataset such as herbarium,
imagenet22k and inaturalist21. The results are shown in Table 7. Experimentally, we do not
directly choose the best-performed pre-training model. Instead, we use ensemble techniques
to combine them. We find that ensemble will produce a consistent improvement compared
with single model. Even combining the best-performed single model with other slightly poorer-
performed models will not affect the conclusion.
3.3. Other Attempts
ConvNext. In addition to MetaFormer, We also train ConvNext. The results are listed in Table 8.
The experiments in ConvNext is not fully explored compared to MetaFormer. Although the
results of ConvNext are inferior to MetaFormer, we still add it to the model ensemble process
and the performance is also improved.
Table 9
Mean π 1 score on public/private test set with pseudo label, ConvNext-large and MetaFormer-2
as backbone.
backbone batch size accumulate steps epochs public mean π 1 private mean π 1
ConvNext-large 24 6 80 80.65% 76.61%
MetaFormer-2 24 8 80 82.45% 77.93%
Table 10
Single modelβs mean π 1 score on public/private test set with different test time augmentation.
test time augmentation public mean π 1 private mean π 1
center crop / five crop 81.66% / 81.76% 77.94% / 78.25%
center crop / five crop 81.67% / 81.63% 77.62% / 77.69%
five crop / multi scale & ten crop 80.46% / 80.20% 77.02% / 77.31%
Table 11
Ensemble modelβs mean π 1 score on public/private test set with different test time augmentation.
test time augmentation public mean π 1 private mean π 1
center crop / multi scale & ten crops 83.20% / 83.26% 79.51% / 79.38%
Table 12
The effectiveness of post process for tail categories recognition and open-set categories recogni-
tion.
ensemble and post process number of open-set samples public mean π 1 private mean π 1
average ensemble (v1) βΌ1000 83.26% 79.38%
average ensemble (v2) βΌ1500 83.50% 79.60%
average ensemble (v3) βΌ1500 83.65% 79.79%
average ensemble (v4) βΌ1500 83.78% 80.43%
Pseudo label. After training models with various settings, we use model ensemble to get the
best model currently, and take the model predictions on test samples as their label. We select top
βΌ 50% test samples by their confidence score. We trained MetaFormer-2 and ConvNext-large
with train+val+pseudo, the results are listed in Table 9.
3.4. Test Time Augmentation and Post Process
Test Time Augmentation. For test time augmentation, we use center crop, five crop and multi
scale & ten crop during test phase. The effects of test time augmentation(TTA) are shown in
Table 10 and Table 11. It should be noted that the mean π 1 score on public test set is out of
accord with private test set in some experiments, and it is hard to decide which TTA is better only
based on public score, in consideration of robustness, we have chosen multi scale & ten crop
based on the public mean π 1 in Table 11.
Post Process. For short, we name the different version of ensemble and post process as v1 (initial
version), v2 (v1 + proper open-set threshold), v3 (v2 + models trained with pseudo label) and
v4 (v3 + post process for tail categories).
For open-set recognition, we intuitively select samples with lower confidence score as open-set
samples. As shown in Table 12, by increasing open-set sample from βΌ 1000 to βΌ 1500, we
improve mean π 1 score from 83.26% to 83.50% on public test set, from 79.38% to 79.60% on
private test set. It demonstrates the effectiveness to select a proper open-set threshold. Compare
the average ensemble (v3) with average ensemble (v2), v3 improves the ensemble performance,
the only difference between them is that v3 contains the models trained with pseudo label. It
is acknowledged that the tail categories tend to have lower confidence score compared to head
categories, so the tail categories are easier to be misclassified as head categories or wrongly
identified as open-set categories. As shown in Table 12, with our post process for tail categories
applied on average ensemble (v3), which termed as average ensemble (v4), the mean π 1 score
improves a lot on both public test set and private test set.
4. Conclusion
In this paper, we introduce our solution for FungiCLEF 2022 competition. To solve this chal-
lenging fine-grained, open-set problem, we try a bunch of techniques, such as different network
baseline, hyper-parameters tuning, modern training techniques, loss for long tail recognition
and specially designed post process. With these endeavours we achieved 1st place among the
participators. The experimental results show the progressive process for single model, and the
effectiveness of test time augmentation and post process for tail categories. For future work,
it is valuable to study the method that fuse meta-information and visual information for Fine-
Grained Visual Classification, and the problem of distinguish between tail categories and open-set
categories is also worth exploring.
References
[1] Q. Diao, Y. Jiang, B. Wen, J. Sun, Z. Yuan, Metaformer: A unified meta framework for fine-grained
recognition (2022). arXiv:2203.02751.
[2] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, S. Xie, A convnet for the 2020s, Proceedings
of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022).
[3] J. Wang, W. Zhang, Y. Zang, Y. Cao, J. Pang, T. Gong, K. Chen, Z. Liu, C. C. Loy, D. Lin, Seesaw loss
for long-tailed instance segmentation, in: Proceedings of the IEEE/CVF International Conference on
Computer Vision, 2021, pp. 9695β9704.
[4] L. Picek, M. Ε ulc, J. Heilmann-Clausen, J. Matas, Overview of FungiCLEF 2022: Fungi recognition
as an open set classification problem, in: Working Notes of CLEF 2022 - Conference and Labs of
the Evaluation Forum, 2022.
[5] A. Joly, H. GoΓ«au, S. Kahl, L. Picek, T. Lorieul, E. Cole, B. Deneu, M. Servajean, A. Durso, I. Bolon,
et al., Lifeclef 2022 teaser: An evaluation of machine-learning based species identification and
species distribution prediction, in: European Conference on Information Retrieval, Springer, 2022,
pp. 390β399.
[6] A. Joly, H. GoΓ«au, S. Kahl, L. Picek, T. Lorieul, E. Cole, B. Deneu, M. Servajean, A. Durso,
H. Glotin, R. PlanquΓ©, W.-P. Vellinga, A. Navine, H. Klinck, T. Denton, I. Eggel, P. Bonnet, M. Ε ulc,
M. Hruz, Overview of lifeclef 2022: an evaluation of machine-learning based species identification
and species distribution prediction, in: International Conference of the Cross-Language Evaluation
Forum for European Languages, Springer, 2022.
[7] L. Picek, M. Šulc, J. Matas, J. Heilmann-Clausen, T. S. Jeppesen, T. Læssøe, T. Frøslev, Danish
fungi 2020 - not just another image recognition dataset (2021). arXiv:2103.10107.
[8] W. Ge, X. Lin, Y. Yu, Weakly supervised complementary parts models for fine-grained image
classification from the bottom up, in: Proceedings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition, 2019, pp. 3034β3043.
[9] C. Liu, H. Xie, Z.-J. Zha, L. Ma, L. Yu, Y. Zhang, Filtration and distillation: Enhancing region
attention for fine-grained visual categorization, in: Proceedings of the AAAI Conference on Artificial
Intelligence, volume 34, 2020, pp. 11555β11562.
[10] Z. Yang, T. Luo, D. Wang, Z. Hu, J. Gao, L. Wang, Learning to navigate for fine-grained classification,
in: Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 420β435.
[11] J. He, J.-N. Chen, S. Liu, A. Kortylewski, C. Yang, Y. Bai, C. Wang, A. Yuille, Transfg: A transformer
architecture for fine-grained recognition (2021). arXiv:2103.07976.
[12] Y. Gao, X. Han, X. Wang, W. Huang, M. Scott, Channel interaction networks for fine-grained image
categorization, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, 2020,
pp. 10818β10825.
[13] X. He, Y. Peng, Fine-grained image classification via combining vision and language, in: Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 5994β6002.
[14] G. Chu, B. Potetz, W. Wang, A. Howard, Y. Song, F. Brucher, T. Leung, H. Adam, Geo-aware
networks for fine-grained recognition, in: Proceedings of the IEEE/CVF International Conference
on Computer Vision Workshops, 2019, pp. 0β0.
[15] O. Mac Aodha, E. Cole, P. Perona, Presence-only geographical priors for fine-grained image
classification, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019,
pp. 9596β9606.
[16] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani,
M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words:
Transformers for image recognition at scale, in: 9th International Conference on Learning Rep-
resentations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, OpenReview.net, 2021. URL:
https://openreview.net/forum?id=YicbFdNTTy.
[17] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, B. Guo, Swin transformer: Hierarchical
vision transformer using shifted windows, in: Proceedings of the IEEE/CVF International Conference
on Computer Vision (ICCV), 2021.
[18] X. Dong, J. Bao, D. Chen, W. Zhang, N. Yu, L. Yuan, D. Chen, B. Guo, Cswin transformer: A
general vision transformer backbone with cross-shaped windows, 2021. arXiv:2107.00652.
[19] T. Wang, Y. Zhu, C. Zhao, W. Zeng, J. Wang, M. Tang, Adaptive class suppression loss for long-tail
object detection, 2021.
[20] J. Tan, C. Wang, B. Li, Q. Li, W. Ouyang, C. Yin, J. Yan, Equalization loss for long-tailed object
recognition, in: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR
2020, Seattle, WA, USA, June 13-19, 2020, Computer Vision Foundation / IEEE, 2020, pp. 11659β
11668. URL: https://openaccess.thecvf.com/content_CVPR_2020/html/Tan_Equalization_Loss_for_
Long-Tailed_Object_Recognition_CVPR_2020_paper.html. doi:10.1109/CVPR42600.2020.
01168.
[21] S. Vaze, K. Han, A. Vedaldi, A. Zisserman, Open-set recognition: a good closed-set classifier is all
you need?, in: International Conference on Learning Representations, 2022.
[22] H. Zhang, M. CissΓ©, Y. N. Dauphin, D. Lopez-Paz, mixup: Beyond empirical risk minimization,
in: 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada,
April 30 - May 3, 2018, Conference Track Proceedings, OpenReview.net, 2018. URL: https://
openreview.net/forum?id=r1Ddp1-Rb.
[23] R. MΓΌller, S. Kornblith, G. Hinton, When does label smoothing help?, in: Proceedings of the 33rd
International Conference on Neural Information Processing Systems, Curran Associates Inc., Red
Hook, NY, USA, 2019.