Despite the recent success of deep learning architectures, person
re-identification (ReID) remains a challenging problem in real-word
applications. Several unsupervised single-target domain adaptation (STDA)
methods have recently been proposed to limit the decline in ReID accuracy
caused by the domain shift that typically occurs between source and target
video data. Given the multimodal nature of person ReID data (due to variations
across camera viewpoints and capture conditions), training a common CNN
backbone to address domain shifts across multiple target domains, can provide
an efficient solution for real-time ReID applications. Although multi-target
domain adaptation (MTDA) has not been widely addressed in the ReID literature,
a straightforward approach consists in blending different target datasets, and
performing STDA on the mixture to train a common CNN. However, this approach
may lead to poor generalization, especially when blending a growing number of
distinct target domains to train a smaller CNN.
To alleviate this problem, we introduce a new MTDA method based on knowledge
distillation (KD-ReID) that is suitable for real-time person ReID applications.
Our method adapts a common lightweight student backbone CNN over the target
domains by alternatively distilling from multiple specialized teacher CNNs,
each one adapted on data from a specific target domain. Extensive experiments
conducted on several challenging person ReID datasets indicate that our
approach outperforms state-of-art methods for MTDA, including blending methods,
particularly when training a compact CNN backbone like OSNet. Results suggest
that our flexible MTDA approach can be employed to design cost-effective ReID
systems for real-time video surveillance applications.
KNOWLEDGE DISTILLATION FOR MULTI-TARGET DOMAIN ADAPTATION IN
マルチターゲット領域適応のための知識蒸留
0.37
REAL-TIME PERSON RE-IDENTIFICATION
リアルタイムパーソン再識別
0.45
Félix Remigereau, Djebril Mekhazni, Sajjad Abdoli, Le Thanh Nguyen-Meidine, Rafael M. O. Cruz and Eric Granger
Félix Remigereau, Djebril Mekhazni, Sajjad Abdoli, Le Thanh Nguyen-Meidine, Rafael M. O. Cruz, Eric Granger
0.48
Laboratoire d’Imagerie, de Vision et d’Intelligence Artificielle (LIVIA)
視覚と知的能力(livia)の労働力
0.32
Dept. of Systems Engineering, École de technologie supérieure, Montreal, Canada
カナダ・モントリオール工科大学システム工学科
0.36
ABSTRACT Despite the recent success of deep learning architectures, person re-identification (ReID) remains a challenging problem in real-word applications.
Several unsupervised single-target domain adaptation (STDA) methods have recently been proposed to limit the decline in ReID accuracy caused by the domain shift that typically occurs between source and target video data.
Given the multimodal nature of person ReID data (due to variations across camera viewpoints and capture conditions), training a common CNN backbone to address domain shifts across multiple target domains, can provide an efficient solution for real-time ReID applications.
Although multi-target domain adaptation (MTDA) has not been widely addressed in the ReID literature, a straightforward approach consists in blending different target datasets, and performing STDA on the mixture to train a common CNN.
To alleviate this problem, we introduce a new MTDA method based on knowledge distillation (KD-ReID) that is suitable for real-time person ReID applications.
Our method adapts a common lightweight student backbone CNN over the target domains by alternatively distilling from multiple specialized teacher CNNs, each one adapted on data from a specific target domain.
Extensive experiments1 conducted on several challenging person ReID datasets indicate that our approach outperforms state-of-art methods for MTDA, including blending methods, particularly when training a compact CNN backbone like OSNet.
ing from sports analytics to video surveillance [1].
スポーツ分析からビデオサーベイランスまで[1]。
0.54
State-ofthe-art approaches for person ReID are typically implemented with a deep learning (DL) model, e g , deep Siamese networks, trained through metric learning to provide global appearance features.
ReIDの最先端のアプローチは一般的に、ディープ・ラーニング(DL)モデル、例えばディープ・シームズ・ネットワークで実装され、メトリック・ラーニングを通じて訓練され、グローバルな外観機能を提供する。 訳抜け防止モード: state - ofthe - art approach for person reidは通常、ディープラーニング(dl)モデルで実装される。 例えば、deep siamese networksは、メトリック学習を通じてグローバルな外観機能を提供するように訓練されている。
0.60
CNN backbones are often used to learn an embedding, where similar image pairs (with the same identity) are close to each other, and dissimilar image pairs (with different identities) are distant from each other.
They also suffer from poor generalization in the presence of domain shift, where the distribution of original video data captures from the source domain diverges w.r.t data from the operational target domain.
Domain shifts are introduced by variations in capture conditions (i.e., pose and illumination), camera settings and viewpoints from which a person is observed, and lead to considerable changes in the distribution of images in different datasets [3, 4].
Given the cost of data annotation, several unsupervised STDA methods have recently been proposed to limit the decline of ReID accuracy caused by domain shift.
These methods seek to adapt CNNs trained with annotated source video data to perform well in a target domain by leveraging unlabeled data captured from that domain.
To learn a discriminant domaininvariant feature representation from source and target data, STDA methods rely on, e g , discrepancy-based or adversarial approaches [5, 3, 6, 7, 8, 9, 10, 11, 12, 13].
Existing STDA techniques can be extended to MTDA by either adapting multiple models, one for each target domain, or by blending data from the multiple target domains, and then applying STDA on the mixture of target data.
For image classification, an adversarial MTDA approach based on blending transfer was recently proposed where all targets are viewed as one domain [15], target-specific representations were concatenated to deploy a common model [16].
Nevertheless, these approaches are either too complex it requires one model per target domain, or generalize poorly on distinct target domains, particularly when adapting a smaller
common CNN backbone on a growing number of targets.
ターゲット数の増加に対する共通cnnバックボーン。
0.52
In [6], MTDA is performed by distilling information from targetspecific teachers to a student model (deployed for testing), significantly reducing system complexity.
In person ReID, video cameras in a distributed network may correspond to different target domains (defined by viewpoints and capture conditions), and the CNN backbone should therefore be adapted to generalize well across multiple target domains.
More recently, Tian et al [17] introduced an MTDA method for ReID.
最近では、tian et al [17] が reid の mtda メソッドを導入した。
0.66
Inspired by the importance of intra-domain camera style, they propose a Camera Identity-guided Distribution Consistency method based on distribution consistency and discriminative embedding losses.
The first loss aligns image style of source and targets through a generative approach, while the second predicts the camera for each identity to decrease distances for each individual across cameras.
However, for real-time person ReID, specialized MTDA techniques are required to train smaller cost-effective CNN backbones that can address domain shifts across multiple data distributions.
In this paper, we advocate for MTDA methods based on knowledge distillation (KD) [18, 14] to provide a better trade-off between CNN accuracy and efficiency.
It was applied in a semi-supervised learning scenario and distills knowledge from multiple source models (teachers) to a target model (student) by aligning studentteacher similarity matrices.
Let xs ∈ X s be the set of samples from the source domain, and ys ∈ Y be their corresponding labels, where X s is the source domain feature space, and Y is the source label space.
xs ∈ X s をソース領域からのサンプルの集合とし、ys ∈ Y を対応するラベルとし、X s はソース領域特徴空間、Y はソース領域特徴空間とする。 訳抜け防止モード: xs ∈ X s をソース領域からのサンプルの集合とする。 そして ys ∈ Y は対応するラベルであり、X s は元領域の特徴空間である。 そして Y はソースラベル空間である。
0.77
Given the target domains feature space X t, let xt = T} ∈ X t be the set of samples for T unlabeled {xt target domain datasets.
対象領域の特徴空間 X t に対し、xt = T} ∈ X t を T の未ラベル {xt の対象領域データセットのサンプル集合とする。
0.75
We define the common CNN backbone as Θ, and a set of target-specific CNN backbones, each one adapted to each target domain, as Φ = {Φ1, Φ2, ..., ΦT}.
As illustrated in Figure 1 (a), a straightforward MTDA method for adapting a common CNN backbone Θ consists in blending the data xt i of all of the targets for i = 1, 2, ..., T , and then applying a STDA method on the resulting dataset xt.
図1(a)に示すように、共通のCNNのバックボーンを適応するための単純なMTDA法は、i = 1, 2, ..., T のすべてのターゲットのデータ xt i を混合し、結果のデータセット xt にSTDA法を適用する。 訳抜け防止モード: 図1(a)に示すように 共通のCNNバックボーンを適応するための単純なMTDA法は、i = 1 のすべてのターゲットのデータ xt i をブレンドすることで構成される。 2, ..., T, そして結果のデータセット xt に STDA メソッドを適用する。
0.80
(a) (b) Fig. 1.
(a) (b) 図1。
0.38
Overview of MTDA methods. (a) Blending: target domain datasets are combined to form a dataset, and the common CNN is adapted using a STDA method.
As illustrated in Figure 1 (b), our proposed method distills knowledge learned by multiple specialized CNN backbones Φi (teacher models), each one adapted on data xt i from a specific target domain, for i = 1, 2, ..., T , using a STDA method.
図1(b)に示すように、提案手法では、特定の対象領域からのデータxt i、例えばi = 1, 2, ..., t に対してstda法を用いて、複数のcnnバックボーン φi (teacher model) で学習した知識を蒸留する。 訳抜け防止モード: 図1(b)に示すように、提案手法は、複数の専門的なCNNバックボーン(教師モデル)によって学習された知識を抽出する。 それぞれが特定のターゲットドメインからのデータ xt i に適合します。 i = 1, 2, ..., T, STDAメソッドを使用する。
0.79
In particular, the KD-ReID method consists of three steps.
特に、KD-ReID法は3つのステップからなる。
0.67
(1) Pre-training on source data: All the CNN backbones (teachers Φi and student Θ models) undergo supervised pre-training on the labeled source domain dataset xs.
This process should provide discriminant networks for person ReID on the source data.
このプロセスは、ソースデータ上の人物ReIDのための識別ネットワークを提供する。
0.66
Any appropriate supervised loss functions for metric learning can be applied, including softmax cross-entropy [20], and hard samples mining triplet loss [21].
(2) STDA of teacher CNNs: The teacher models Φi are adapted to their respective target domains using some STDA method with source xs and target xt i datasets, for i = 1, 2, ..., T .
2)教師CNNのSTDA:i = 1, 2, ..., T のソース xs とターゲット xt i データセットを持つ STDA メソッドを用いて,教師モデル .i をそれぞれのターゲットドメインに適応させる。
0.82
Any STDA method can be applied, including the D-MMD [3], and SPCL [22] methods.
D-MMD[3]やSPCL[22]メソッドを含む任意のSTDAメソッドを適用することができる。
0.78
(3) KD to student CNN: Knowledge from every teacher model Φi is distilled to the common student model Θ based on target datasets xt i, for i = 1, 2, ..., T , and on the KD loss function, LKD (described below).
(3) 学生CNNへのKD: すべての教師モデルからの知識は、目標データセット xt i, for i = 1, 2, ..., T, and on the KD loss function LKD(後述)に基づいて、共通の学生モデルに蒸留される。
0.82
The KD-ReID method is summarized in Algorithm 1.
KD-ReID法はアルゴリズム1で要約される。
0.69
Once the set of teacher models Φ are adapted to respective target domains, KD is proposed to progressively integrate their knowledge into the common student model Θ.
Based on the self-similarity matrices of feature vectors from Φ and Θ, our
φ と θ による特徴ベクトルの自己相似性行列に基づく。
0.72
英語（論文から抽出）
日本語訳
スコア
Algorithm 1 KD-ReID method.
アルゴリズム1 KD-ReID法
0.79
Require: labeled source datasets: xs, unlabeled target data: xt = {xt 1) Pre-train teacher models, Φ = {Φ1, Φ2, ..., ΦT}, and student model, Θ, on the source labeled data xs for i = 1, 2, ...T do for each mini-batches Bs ⊂ xs and Bt ⊂ xt
必須事項: ラベル付きソースデータセット: xs, unlabeled target data: xt = {xt 1) 事前トレーニングされた教師モデル φ = { φ1, φ2, ..., φt} と学生モデル θ, on the source labeled data xs for i = 1, 2, ...t do for each mini-batches bs , xs and bt , xt
0.88
2, ..., xt
2... ... xt
0.37
T}, 1, xt
t} である。 1,xt
0.46
i do 2) STDA: adapt teacher Φi on source and target data
私は 2)STDA:ソースデータとターゲットデータに基づく教員のi適応
0.74
for each mini-batch Bt ⊂ xt
各ミニバッチ bt , xt について
0.56
i do 3) KD: adapt student Θ using target data and LKD
私は 3)KD: 対象データとLKDを用いた適応型学習者
0.71
end for end for for i = 1, 2, ...T do
終わりだ i = 1, 2, ... の場合の終了
0.65
end for end for proposed KD loss LKD aligns student-teacher representations.
終わりだ 終わりだ 提案したKD損失LKDは生徒と教師の表現を一致させる。
0.53
Let fn = Φi(xn) be the feature vector output from the fully connected layer of Φi in response to sample xn.
The self-similarity matrix denoted as A ∈ RN×N is defined as aj,k = (cid:104)fj, fk(cid:105), where (cid:104)·,·(cid:105) is the cosine distance between the feature vectors and fj, fk ∈ F .
A ∈ RN×N と表される自己相似行列は aj,k = (cid:104)fj, fk(cid:105), ここで (cid:104)·,·(cid:105) は特徴ベクトルと fj, fk ∈ F の間の余弦距離である。
0.83
Similarity matrices As and At are computed based on student Fs and teacher Ft feature matrices, respectively.
類似度行列AsとAtはそれぞれ、学生Fsと教師Ftの特徴行列に基づいて計算される。
0.71
Our KD loss LKD is then computed as: LKD = ||As − At||F , where ||·,·||F is the Frobenius norm matrix.
Optimizing Θ parameters with LKD allows the student model to produce a distance matrix aligned to the teacher model.
lkd で θ パラメータを最適化することで,教師モデルに沿った距離行列を生成することができる。
0.77
We only apply this loss on target samples xt.
この損失はターゲットサンプル xt にのみ適用される。
0.82
It allows tuning Θ parameters, while Φ parameters are fixed.
パラメーターをチューニングできるが、パラメーターは固定されている。
0.73
For each mini-batch Bt from a given target domain i, we select only the corresponding teacher model Φi.
対象領域 i からの各ミニバッチ Bt に対して、対応する教師モデル .i のみを選択する。
0.74
The order of targets for KD is selected randomly, and it changes at every epoch.
KDのターゲットの順序はランダムに選択され、各エポックごとに変化する。
0.70
An important benefit of the KD-ReID approach is its versatility.
KD-ReIDアプローチの重要な利点は、その汎用性である。
0.61
It is independent of the STDA methods used to adapt each Φi.
STDA法とは独立であり、それぞれを順応するために用いられる。
0.64
The choice of CNN backbones Φi, and its STDA methods may differ, allowing to select the best configuration to address the specific challenges (i.e., domain shift) faced by each target domain.
Considering that videos are captured in indoor/outdoor conditions, during different days, weather conditions, and times of the day, it certainly represents a more challenging and realistic dataset in comparison to other person ReID datasets.
MSMT17 is always used as the source dataset to pre-train the CNNs.
MSMT17は、常にCNNを事前トレーニングするためのソースデータセットとして使用される。
0.57
Considering the diversity of capture conditions, there is a significant domain shift between samples from these datasets (results not shown in this paper).
Experimental Setting: Resnet50 [27] is used to implement the target-specific CNN backbones (teachers), while Osnet_x0_25 [28] implements the common CNN backbone (deployed for testing).
Training is conducted until the model performance improves by less than 0.5% average mAP over five consecutive epochs.
トレーニングは、モデルの性能が5回連続で平均mAPを0.5%以下に改善するまで実施される。
0.74
Results and Discussion: Table 3 compares the performance of the proposed KD-ReID to the baseline methods in terms of mean Average Precision (mAP), rank-1 accuracy, number of parameters, and FLOPs.
結果と考察:表3は提案したKD-ReIDの性能を平均精度(mAP)、ランク1の精度、パラメータ数、FLOPの基準値と比較する。 訳抜け防止モード: 結果と考察 : 表3では,提案したKD-ReIDと平均精度(mAP)の基準値との比較を行った。 rank-1 accuracy , number of parameters , and FLOPs 。
0.83
As expected, the lower-bound accuracy on target datasets is low due to the domain shift between the source and target datasets.
Table 1. Performance of MTDA methods when MSMT17 is used as the source dataset, and Market1501, DukeMTMC, and CUHK03 are used as target datasets (T = 3 targets), with 2 STDA techniques – D-MMD and SPCL.
The lower bound performance is obtained through supervised training of Osnet_x0_25 on the labeled source dataset only, and then evaluation on the 3 target datasets.
For the upper bound, supervised training of Osnet_x0_25 is performed on the blended target datasets.
上界では、混合ターゲットデータセット上でOsnet_x0_25の教師付きトレーニングを行う。
0.63
"*": For the "One Model / Target" methods, the Resnet50 (target-specific) performance is presented as an upper bound that should not be directly compared with other MTDA techniques.
"*": "One Model / Target" メソッドでは、Resnet50(ターゲット固有の)パフォーマンスは、他のMTDA技術と直接比較するべきではない上限として提示される。
0.75
FLOPs are related to the extraction features for one sample.
FLOPは1つのサンプルの抽出機能と関連している。
0.69
and the labeled data from multiple different source domains to adapt the common CNN backbone.
そして、複数の異なるソースドメインからのラベル付きデータを共通のCNNバックボーンに適合させる。
0.67
In practice, this scenario is restrictive for ReID, given the need to collect datasets and annotate samples.
Second, Tian et al [17] recently proposed the only MTDA method in literature for ReID.
第2に、Tian et al [17]はReIDの文献における唯一のMTDA法を提案した。
0.66
Our method significantly outperforms this approach with an improvement of 18.3% in mAP, and 6.2% in rank-1 on average, even on a much smaller CNN architecture (by a ratio close to 60).
In contrast, our KD-ReID method divides the problem into simpler STDA problems, that may be solved using any combination of suitable SOTA methods, rather than resolving the complex MTDA problem.
The experiments are conducted using three variations of the Osnet CNN: osnet_x1_0, osnet_x0_5, and osnet_x0_25, having about 4.0M, 1.0M, and 0.4M parameters, respectively.
We consider multiple models for [18] since it is a MSDA method, where T is the number of targets.
T が対象数である MSDA メソッドであるため,[18] に対する複数のモデルを考える。
0.72
ReID always outperforms the blending technique and remains significantly more effective for adapting lightweight models, thus more suitable for real-time applications.
The common CNN is Osnet, while teachers are Resnet50s, and D-MMD is used for STDA.
一般的なCNNはOsnetであり、教師はResnet50であり、D-MMDはSTDAに使われる。
0.69
4. CONCLUSION
4.コンキュレーション
0.72
In this paper, KD-ReID is introduced as a versatile MTDA method for real-time person ReID.
本稿では,リアルタイム人物ReIDのための汎用MTDA手法としてKD-ReIDを提案する。
0.66
To adapt each teachers model, KD-ReID allows selecting the STDA method and CNN individually, and then knowledge of all teacher models is distilled to a common lightweight CNN (student).
Our experiments conducted on challenging person ReID datasets indicate that student models adapted through KD-ReID outperform blending methods, and generalize well on all targets at the same time, even when adapting a small common CNN backbone on a growing number of targets.
5. REFERENCES [1] Y. He, X. Wei, X. Hong, W. Shi, and Y. Gong, “Multitarget multi-camera tracking by tracklet-to-target assignment,” IEEE Trans.
5.参考文献 [1] Y. He, X. Wei, X. Hong, W. Shi, Y. Gong, “Multitarget multi-camera tracking by tracklet-to-target assignment”, IEEE Trans。
0.57
Image Processing, vol.
画像処理、vol.1。
0.62
29, pp. 5191– 5205, 2020.
pp. 5191– 5205, 2020。
0.67
[2] M. Boudiaf, J. Rony, I.M. Ziko, E. Granger, M. Pedersoli, P. Piantanida, and I. Ben Ayed, “A unifying mutual information view of metric learning: cross-entropy vs. pairwise losses,” in ECCV 2020.
[2] m. boudiaf, j. rony, i.m. ziko, e. granger, m. pedersoli, p. piantanida, i. ben ayed, “a unified mutual information view of metric learning: cross-entropy vs. pairwise loss” (eccv 2020) の著者である。 訳抜け防止モード: [2]M.ブーダフ、J.ロニー、I.M.ジコ E. Granger, M. Pedersoli, P. Piantanida, I. Ben Ayed 「メートル法学習の統一的相互情報視点 : クロス-エントロピー対対対損失」 ECCV 2020に登場。
0.79
[3] D. Mekhazni, A. Bhuiyan, G. Ekladious, and E. Granger, “Unsupervised DA in the dissimilarity space for person reID,” in ECCV 2020.
[3] d. mekhazni, a. bhuiyan, g. ekladious, e. granger, “unsupervised da in the different similarity space for person reid” (eccv 2020)で発表された。
0.59
[4] S. Xuan and S. Zhang, “Intra-inter camera similarity for
4]S.XuanとS.Zhang:「イントラインターカメラの類似性」
0.67
unsupervised person reID,” arXiv 2021.
とarxiv 2021は言った。
0.31
[5] L. Wang and K Yoon,
5]l・ワンとk・ユン
0.53
“Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks,” IEEE Trans.
[6] L.T. Nguyen-Meidine, A. Belal, M. Kiran, J. Dolz, L. A. Blais-Morin, and E. Granger, “Knowledge distillation methods for efficient unsupervised adaptation across multiple domains,” Image and Vision Computing, vol.
6] l.t. nguyen-meidine, a. belal, m. kiran, j. dolz, l. a. blais-morin, e. granger, “複数の領域にわたる効率的な非教師なし適応のための知識蒸留法”。 訳抜け防止モード: [6 ]L.T.Nguyen-Meidine, A. Belal, M. Kiran, J. Dolz, L. A. Blais - Morin, E. Granger 『複数の領域にまたがる効率的な非教師なし適応のための知識蒸留方法』 Image and Vision Computing, vol。
0.72
108, 2021.
108, 2021.
0.43
[7] L.T. Nguyen-Meidine, E. Granger, M. Kiran, J. Dolz, and L.A. Blais-Morin, “Joint progressive knowledge distillation and unsupervised DA,” arXiv 2020.
[7] L.T. Nguyen-Meidine, E. Granger, M. Kiran, J. Dolz, L.A. Blais-Morin, “Joint progressive knowledge distillation and unsupervised DA”, arXiv 2020。 訳抜け防止モード: [7 ]L.T.Nguyen-Meidine, E. Granger, M. Kiran, J. Dolz, and L.A. Blais - Morin, “ジョイントプログレッシブな知識蒸留と教師なしDA”。 arXiv 2020。
0.69
[8] Z. Liu, Z. Miao, X. Pan, X. Zhan, D. Lin, S. Yu, and
[8]Z.Liu、Z.Miao、X.Pan、X.Zhan、D.Lin、S.Yu、
0.38
B. Gong, “Open compound DA,” in CVPR 2020.
b. gong, “open compound da” in cvpr 2020(英語)
0.67
[9] G. Ekladious, H. Lemoine, E. Granger, K. Kamali, and S. Moudache, “Dual-triplet metric learning for unsupervised DA in video-based face recognition,” arXiv 2020.
9]g. ekladious, h. lemoine, e. granger, k. kamali, s. moudache, “ビデオベースの顔認識で教師なしのdaのためのデュアルトリップレットメトリックラーニング”。 訳抜け防止モード: [9 ]G. Ekladious, H. Lemoine, E. Granger, K. Kamali, S. Moudache両氏は,“デュアル – 教師なしDAの3倍のメトリクス学習 – 顔認識に基づく”と述べている。 arXiv 2020。
0.63
[10] L.T. Nguyen-Meidine, M. Kiran, M. Pedersoli, J. Dolz, L.A. Blais-Morin, and E. Granger, “Incremental multitarget DA for object detection with efficient domain transfer,” arXiv 2021.
10] l.t. nguyen-meidine, m. kiran, m. pedersoli, j. dolz, l.a. blais-morin, e. granger, "効率的なドメイン転送によるオブジェクト検出のためのインクリメンタルマルチターゲットda", arxiv 2021。 訳抜け防止モード: [10 ]L.T.Nguyen-Meidine,M .Kiran,M.Pedersoli J. Dolz, L.A. Blais - Morin, E. Granger 「効率的なドメイン転送によるオブジェクト検出のためのインクリメンタルマルチターゲットDA」 arXiv 2021。
0.73
[11] Y. Ge, Z. Li, H. Zhao, G. Yin, S. Yi, X. Wang, and H. Li, “Fd-gan: Pose-guided feature distilling gan for robust person reID,” in NeurIPS 2018.
He11] Y. Ge, Z. Li, H. Zhao, G. Yin, S. Yi, X. Wang, H. Li, “Fd-gan: Pose-guided feature distilling gan for robust person reID”. NeurIPS 2018. 訳抜け防止モード: [11]Y.Ge,Z.Li,H.Zhao, G.Yin,S.Yi,X.Wang,H. Li 「Fd-gan : Pose- Guided feature distilling gan for robust person reID」 2018年のNeurIPS。
0.79
[12] S. Khurana, N. Moritz, T. Hori, and J. Le Roux, “Unsupervised domain adaptation for speech recognition via uncertainty driven self-training,” in ICASSP 2021.
12] S. Khurana, N. Moritz, T. Hori, J. Le Roux, “Unsupervised domain adapt for speech Recognition through uncertainty driven self-training” in ICASSP 2021。 訳抜け防止モード: [12 ]S.Khurana,N. Moritz,T.Hori とJ. Le Roux氏は語る。「不確実性駆動型自己学習による音声認識のための教師なしドメイン適応」。 ICASSP 2021。
0.68
[13] Z. Allen-Zhu and Y. Li, “Towards understanding ensemble, knowledge distillation and self-distillation in deep learning,” arXiv 2020.
Z. Allen-Zhu 氏と Y. Li 氏は,“深層学習におけるアンサンブル,知識蒸留,自己蒸留を理解する” と arXiv 2020 だ。
0.75
[14] L.T. Nguyen-Meidine, A. Belal, M. Kiran, J. Dolz, L.A. Blais-Morin, and E. Granger, “Unsupervised multi-target DA through knowledge distillation,” in WACV 2021.
Nguyen-Meidine, A. Belal, M. Kiran, J. Dolz, L.A. Blais-Morin, E. Granger, “Unsupervised multi-target DA through knowledge distillation” in WACV 2021。 訳抜け防止モード: [14 ]L.T.Nguyen-Meidine, A. Belal, M. Kiran, J. Dolz, L.A. Blais - Morin, E. Granger WACV2021における「知識蒸留による非監督マルチターゲットDA」
0.72
[15] Z. Chen, J. Zhuang, X. Liang, and L. Lin, “Blendingtarget domain adaptation by adversarial meta-adaptation networks,” in CVPR 2019.
15] z. chen, j. zhuang, x. liang, l. lin, “blendingtarget domain adapt by adversarial meta-adaptation networks” cvpr 2019で発表された。
0.78
[16] B. Gholami, P. Sahu, O. Rudovic, K. Bousmalis, and V. Pavlovic, “Unsupervised multi-target domain adaptation: An information theoretic approach,” IEEE Trans.
[16] b. gholami, p. sahu, o. rudovic, k. bousmalis, v. pavlovic, “教師なしのマルチターゲットドメイン適応: 情報理論的なアプローチ”、ieee transは述べている。
0.62
Image Processing, vol.
画像処理、vol.1。
0.62
29, pp. 3993–4002, 2020.
29, pp. 3993-4002, 2020。
0.84
[17] J. Tian, Q. Tang, R. Li, Z. Teng, B. Zhang, and J. Fan, “A camera identity-guided distribution consistency method for unsupervised multi-target domain person reID,” ACM Trans.
[17]j. tian, q. tang, r. li, z. teng, b. zhang, j. fan, “教師なしマルチターゲットドメインのreidに対するカメラidによる分散一貫性の方法”、とacm transは述べている。 訳抜け防止モード: [17 ]J. Tian, Q. Tang, R. Li, Z. Teng, B. Zhang, J. Fan 氏は,“カメラアイデンティティ - 教師なしマルチターゲットドメインパーソン reID の分散一貫性をガイドする”。 ACMトランス。
0.72
Intelligent Systems and Technology, vol.
Intelligent Systems and Technology, vol. (英語)
0.75
12, no. 4, pp. 1–18, 2021.
12,4, pp. 1-18, 2021。
0.76
[18] A. Wu, W.S. Zheng, X. Guo, and J.H. Lai, “Distilled person reID : Towards a more scalable system,” in CVPR 2019.
18] a. wu, w.s. zheng, x. guo, j.h. lai, “distilled person reid: towards a scalable system” cvpr 2019で紹介された。
0.71
[19] A. Wu, W.S. Zheng, X. Guo, and J.H. Lai, “Distilled person reID : Towards a more scalable system,” in CVPR 2019.
19] a. wu, w.s. zheng, x. guo, j.h. lai, “distilled person reid: towards a scalable system” cvpr 2019で紹介された。
0.71
[20] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” in ICCV 2016.
C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, “Rethinking the Inception Architecture for computer vision” in ICCV 2016. ICCV 2016 訳抜け防止モード: [20]C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, “コンピュータビジョンの開始アーキテクチャを再考する”。 ICCV 2016に登場。
0.87
[21] A. Hermans, L. Beyer, and B. Leibe, “In defense of the
A. Hermans, L. Beyer, and B. Leibe, “In Defense of the。
0.40
triplet loss for person reID,” arXiv 2017.
とarxiv 2017で発表した。
0.23
[22] Y. Ge, F. Zhu, D. Chen, R. Zhao, et al , “Self-paced contrastive learning with hybrid memory for domain adaptive object re-id,” NeurIPS 2020.
22] y. ge, f. zhu, d. chen, r. zhao, et al, “ドメイン適応オブジェクト再識別のためのハイブリッドメモリによる自己ペースのコントラスト学習”。
0.70
[23] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang, “Deepreid: Deep filter pairing neural network for person reID,” in CVPR 2014.
[23] Wei Li, Rui Zhao, Tong Xiao, Xiaogang Wang, “Deepreid: Deep Filter pairing Neural Network for person reID”とCVPR 2014で述べている。
0.84
[24] L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian, “Scalable person reID : A benchmark,” in ICCV 2015.
L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, Q. Tian, “Scalable person reID : A benchmark” in ICCV 2015 訳抜け防止モード: 24 ] l. zheng, l. shen, l. tian s. wang, j. wang, q. tian, “スケーラブルな人物リード: ベンチマーク” 2015年、iccvに移籍。
0.66
[25] Z. Zhang, J. Wu, X. Zhang, and C. Zhang, “Multi-target, multi-camera tracking by hierarchical clustering: Recent progress on dukemtmc project,” arXiv 2017.
Z. Zhang, J. Wu, X. Zhang, and C. Zhang, “Multi-target, multi-camera tracking by hierarchical clustering: recent progress on dukemtmc project” arXiv 2017 訳抜け防止モード: 25] z. zhang, j. wu, x. zhang. 階層クラスタリングによるマルチターゲット、マルチカメラトラッキング:dukemtmcプロジェクトにおける最近の進歩” arxiv 2017。
0.60
[26] L. Wei, S. Zhang, W. Gao, and Q. Tian, “Person transfer gan to bridge domain gap for person reID,” in CVPR 2018.
26] l. wei, s. zhang, w. gao, q. tian, “person transfer gan to bridge domain gap for person reid” cvpr 2018で発表された。
0.72
[27] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual
27] he, x. zhang, s. ren, j. sun, 「深い残留」
0.63
learning for image recognition,” in CVPR 2016.
とCVPR 2016で述べている。
0.41
[28] K. Zhou, Y. Yang, A. Cavallaro, and T. Xiang, “Omni-
K. Zhou, Y. Yang, A. Cavallaro, T. Xiang, “Omni-”
0.36
scale feature learning for person reID,” in ICCV 2019.