RT-DETRv3 important words 100

ensure /ɪnˈʃʊr/

【動詞】確実にする、保証する ★★★ 例文：Additionally, we introduce a shared-weight decoder branch for dense positive supervision to ensure more high-quality queries matching each ground truth.（Abstract）

incorporate /ɪnˈkɔːrpəreɪt/

【動詞】取り入れる、組み込む ★★★ 例文：Furthermore, to ensure that there are more high-quality queries matching each ground truth, we incorporate an one-to-many label assignment branch within the decoder.（Related Work）

supervise /ˈsuːpərvaɪz/

【動詞】教師を与える、指導する ★★ 例文：For loss calculation, it uses L1 loss and GIoU loss to supervise box regression, and variable focus loss (VFL) to supervise the learning of the classification task.（Method）

strengthen /ˈstreŋθən/

【動詞】強化する ★★★ 例文：Secondly, to enhance and strengthen supervision of the decoder, we generate multiple object queries (OQ) through the query selection module and apply random masking to perturb the self-attention mechanism, effectively diversifying the distribution of positive query samples.（Related Work）

diversifies /daɪˈvɝːsɪˌfaɪz/ （基本形：diversify）

【動詞】多様化させる ★★★ 例文：This strategy diversifies label assignment for positive samples across multiple query groups, thereby enriching positive supervisions.（Abstract）

enriching /ɪnˈrɪtʃɪŋ/ （基本形：enrich）

【動詞】豊かにする、充実させる ★★ 例文：The introduction of multiple sets of random perturbations diversifies the features of the queries, allowing multiple related queries associated with the same target to have a chance of being assigned as positive sample queries, thereby enriching the supervision information.（Method）

attain /əˈteɪn/

【動詞】達成する、到達する ★★ 例文：Furthermore, RT-DETRv3-R101 can attain an impressive 54.6% AP outperforming YOLOv10-X.（Abstract）

garnered /ˈɡɑːrnərd/ （基本形：garner）

【動詞】集めた、獲得した ★ 例文：In recent years, real-time object detections have garnered significant attention from both researchers and industry professionals due to its efficient inference speed and superior detection accuracy.（Introduction）

sacrificing /ˈsækrɪˌfaɪsɪŋ/ （基本形：sacrifice）

【動詞】犠牲にする ★★ 例文：RT-DETRv2 [19] further enhances the flexibility and practicality of RT-DETR [32] by optimizing the training strategy to improve performance without sacrificing speed, although requires longer training time.（Introduction）

aiming /ˈeɪmɪŋ/ （基本形：aim）

【動詞】目指す ★★ 例文：This series of algorithms has been optimized and improved on the basis of the YOLO series, aiming to improve detection accuracy and speed to meet the needs of practical application scenarios.（Related Work）

diminish /dɪˈmɪnɪʃ/

【動詞】弱める、減少させる ★★ 例文：Initially, the queries interact with each other through the self-attention module to enhance or diminish their feature representations.（Method）

impairing /ɪmˈperɪŋ/ （基本形：impair）

【動詞】損なう、低下させる ★★ 例文：However, the adoption of a one-to-one set matching in the RT-DETR leads to sparse supervision information, ultimately impairing the model's performance.（Method）

yield /jiːld/

【動詞】生み出す、もたらす ★★ 例文：Then, Qi and Ki are multiplied to compute the attention weight, which is further multiplied by Mi and passed through a softmax function to yield the perturbed attention weight.（Method）

designate /ˈdezɪɡneɪt/

【動詞】指定する、呼称する ★ 例文：The loss computation remains consistent with the original detection loss, and we designate Lo2m as the loss function for this branch, with a loss weight of γ.（Method）

reused /riːˈjuːzd/ （基本形：reuse）

【動詞】再利用した ★★ 例文：We reused the RT-DETR [32] decoder structure as the main branch and additionally added three groups of parameter-shared self-attention perturbation branches.（Method）

interact /ˌɪntərˈækt/

【動詞】相互作用する ★★ 例文：Initially, the queries interact with each other through the self-attention module to enhance or diminish their feature representations.（Method）

exhibit /ɪɡˈzɪbɪt/

【動詞】示す、見せる ★★ 例文：As illustrated in Figure 4, we noticed that as the model size increases, RT-DETRv3 tends to exhibit overfitting.（Experiments）

tends /tendz/ （基本形：tend）

【動詞】〜する傾向がある ★★ 例文：As illustrated in Figure 4, we noticed that as the model size increases, RT-DETRv3 tends to exhibit overfitting.（Experiments）

categorized /ˈkætɪɡəˌraɪzd/ （基本形：categorize）

【動詞】分類した ★ 例文：We categorized the models into small, medium, and large scales based on their inference speed.（Experiments）

participate /pɑːrˈtɪsɪpeɪt/

【動詞】参加する、関与する ★★ 例文：To ensure that multiple related queries associated with the same target have the opportunity to participate in positive sample learning, we propose multiple self-attention perturbation modules based on Mask Self-Attention.（Method）

retrieving /rɪˈtriːvɪŋ/ （基本形：retrieve）

【動詞】取得する、取り出す ★★ 例文：Subsequently, each query updates itself by retrieving information from the encoder's output features via the cross-attention module.（Method）

fused /fjuːzd/ （基本形：fuse）

【動詞】融合された ★ 例文：Both OQi and Mi are fed into the Mask Self-Attention module, resulting in the perturbed and fused features.（Method）

boosts /buːsts/ （基本形：boost）

【動詞】押し上げる、高める ★★ 例文：Our approach builds on the RT-DETR [32] framework by incorporating CNN-based and transformer-based one-to-many dense supervision, which not only boosts model performance but also speeds up convergence.（Conclusion）

collaborates /kəˈlæbəˌreɪts/ （基本形：collaborate）

【動詞】協調する、連携する ★★ 例文：Firstly, we introduce a CNN-based auxiliary branch that provides dense supervision that collaborates with the original decoder to enhance the encoder's feature representation.（Abstract）

assigned /əˈsaɪnd/ （基本形：assign）

【動詞】割り当てられた ★★★ 例文：The introduction of multiple sets of random perturbations diversifies the features of the queries, allowing multiple related queries associated with the same target to have a chance of being assigned as positive sample queries, thereby enriching the supervision information.（Method）

noticed /ˈnoʊtɪst/ （基本形：notice）

【動詞】気づいた ★ 例文：As illustrated in Figure 4, we noticed that as the model size increases, RT-DETRv3 tends to exhibit overfitting.（Experiments）

differentiating /ˌdɪfəˈrenʃiˌeɪtɪŋ/ （基本形：differentiate）

【動詞】差別化する、区別する ★★ 例文：These masks are applied to the self-attention module, affecting the correlation between queries and thus differentiating the assignments of positive queries.（Related Work）

switch /swɪtʃ/

【動詞】切り替える ★★ 例文：For the one-to-many matching algorithm, we follow the configuration of the PP-YOLOE head and use the ATSS matching algorithm in the early stage of training, and then switch to the TaskAlign matching algorithm.（Method）

raised /reɪzd/ （基本形：raise）

【動詞】高めた、引き上げた ★★ 例文：Conversely, increasing the number of branches excessively raised the model's learning difficulty without yielding significant performance gains.（Experiments）

branch /bræntʃ/

【名詞】分岐、ブランチ（補助ブランチなど） ★★★ 例文：Firstly, we introduce a CNN-based auxiliary branch that provides dense supervision that collaborates with the original decoder to enhance the encoder's feature representation.（Abstract）

perturbation /ˌpɝːtərˈbeɪʃən/

【名詞】「摂動（せつどう）」や「かく乱（撹乱）」、「動揺」

★★★ 例文：Secondly, to address insufficient decoder training, we propose a novel learning strategy involving self-attention perturbation.（Abstract）

mask /mæsk/

【名詞】マスク ★★★ 例文：Mi represents the perturbation mask corresponding to the i-th set of object queries.（Method）

information /ˌɪnfərˈmeɪʃən/

【名詞】情報 ★★ 例文：However, the adoption of a one-to-one set matching in the RT-DETR leads to sparse supervision information, ultimately impairing the model's performance.（Method）

schedule /ˈskedʒuːl/

【名詞】スケジュール（学習計画） ★★ 例文：We used a 10x (120 epochs) training schedule for smaller backbones (R18, R34) and a 6x (72 epochs) training schedule for larger backbones (R50, R101).（Method）

overfitting /ˌoʊvərˈfɪtɪŋ/

【名詞】過学習 ★★★ 例文：As illustrated in Figure 4, we noticed that as the model size increases, RT-DETRv3 tends to exhibit overfitting.（Experiments）

complexity /kəmˈpleksɪti/

【名詞】複雑さ、計算量 ★★ 例文：However, its high computational complexity significantly limits its practical applications.（Introduction）

distribution /ˌdɪstrɪˈbjuːʃən/

【名詞】分布 ★★★ 例文：Secondly, to enhance and strengthen supervision of the decoder, we generate multiple object queries (OQ) through the query selection module and apply random masking to perturb the self-attention mechanism, effectively diversifying the distribution of positive query samples.（Related Work）

correlation /ˌkɔːrəˈleɪʃən/

【名詞】相関 ★★ 例文：These masks are applied to the self-attention module, affecting the correlation between queries and thus differentiating the assignments of positive queries.（Related Work）

matrix /ˈmeɪtrɪks/

【名詞】行列 ★★ 例文：⊗ denotes matrix multiplication, and ⊙ denotes element-wise multiplication.（Method）

regression /rɪˈɡreʃən/

【名詞】回帰（box regression＝ボックス回帰） ★★★ 例文：For loss calculation, it uses L1 loss and GIoU loss to supervise box regression, and variable focus loss (VFL) to supervise the learning of the classification task.（Method）

consistency /kənˈsɪstənsi/

【名詞】一貫性 ★★ 例文：Specifically, decoder head also use VFL loss to ensure consistency in task definition.（Method）

definition /ˌdefɪˈnɪʃən/

【名詞】定義 ★★ 例文：Specifically, decoder head also use VFL loss to ensure consistency in task definition.（Method）

description /dɪˈskrɪpʃən/

【名詞】説明、記述 ★★ 例文：The following sections provide a detailed description of the modules proposed in this work.（Related Work）

adoption /əˈdɑːpʃən/

【名詞】採用 ★★ 例文：However, the adoption of a one-to-one set matching in the RT-DETR leads to sparse supervision information, ultimately impairing the model's performance.（Method）

opportunity /ˌɑːpərˈtuːnəti/

【名詞】機会 ★★ 例文：To ensure that multiple related queries associated with the same target have the opportunity to participate in positive sample learning, we propose multiple self-attention perturbation modules based on Mask Self-Attention.（Method）

chance /tʃæns/

【名詞】機会、可能性 ★★ 例文：The introduction of multiple sets of random perturbations diversifies the features of the queries, allowing multiple related queries associated with the same target to have a chance of being assigned as positive sample queries, thereby enriching the supervision information.（Method）

phase /feɪz/

【名詞】段階、フェーズ ★★ 例文：Since these modules are only involved during the training phase, they don't affect the inference latency of RT-DETR [32].（Method）

richness /ˈrɪtʃnəs/

【名詞】豊かさ ★ 例文：Reducing the number of branches decreased the richness of the supervision signals, leading to lower performance.（Experiments）

multiplication /ˌmʌltɪplɪˈkeɪʃən/

【名詞】乗算、掛け算 ★★ 例文：⊗ denotes matrix multiplication, and ⊙ denotes element-wise multiplication.（Method）

computation /ˌkɑːmpjuˈteɪʃən/

【名詞】計算 ★★ 例文：The loss computation remains consistent with the original detection loss, and we designate Lo2m as the loss function for this branch, with a loss weight of γ.（Method）

integration /ˌɪntɪˈɡreɪʃən/

【名詞】統合 ★★ 例文：For example, the integration of ViT-CoMer [27] with Co-DETR [33] has achieved state-of-the-art performance on the COCO detection task.（Related Work）

solution /səˈluːʃən/

【名詞】ソリューション、解決策 ★★ 例文：The PP-YOLO series [10,18] is a real-time object detection solution based on the PaddlePaddle framework proposed by Baidu.（Related Work）

basis /ˈbeɪsɪs/

【名詞】基礎、土台（on the basis of＝〜に基づいて） ★★★ 例文：This series of algorithms has been optimized and improved on the basis of the YOLO series, aiming to improve detection accuracy and speed to meet the needs of practical application scenarios.（Related Work）

room /ruːm/

【名詞】余地（room for＝〜の余地） ★★ 例文：However, since we have not yet optimized the overall framework of the RT-DETRv3 detector for lightweight deployment, there is still room for further improving the inference efficiency of RT-DETRv3.（Experiments）

gains /ɡeɪnz/ （基本形：gain）

【名詞】利益、向上 ★★ 例文：Conversely, increasing the number of branches excessively raised the model's learning difficulty without yielding significant performance gains.（Experiments）

researchers /rɪˈsɝːtʃərz/ （基本形：researcher）

【名詞】研究者 ★★ 例文：In recent years, real-time object detections have garnered significant attention from both researchers and industry professionals due to its efficient inference speed and superior detection accuracy.（Introduction）

truth /truːθ/

【名詞】真値（ground truth＝正解ラベル） ★★★ 例文：Additionally, we introduce a shared-weight decoder branch for dense positive supervision to ensure more high-quality queries matching each ground truth.（Abstract）

professionals /prəˈfeʃənəlz/ （基本形：professional）

【名詞】専門家、実務家 ★ 例文：In recent years, real-time object detections have garnered significant attention from both researchers and industry professionals due to its efficient inference speed and superior detection accuracy.（Introduction）

needs /niːdz/ （基本形：need）

【名詞】ニーズ、必要（meet the needs of＝〜の要求を満たす） ★★ 例文：This series of algorithms has been optimized and improved on the basis of the YOLO series, aiming to improve detection accuracy and speed to meet the needs of practical application scenarios.（Related Work）

signals /ˈsɪɡnəlz/ （基本形：signal）

【名詞】信号、シグナル ★★ 例文：Reducing the number of branches decreased the richness of the supervision signals, leading to lower performance.（Experiments）

terms /tɝːmz/ （基本形：term）

【名詞】観点（in terms of＝〜の点で） ★★★ 例文：It clearly demonstrates that our method outperforms them in terms of convergence speed in any schedule and only needs half of the training epochs to achieve the comparable performance.（Experiments）

hierarchical /ˌhaɪəˈrɑːrkɪkəl/

【形容詞】階層的な ★★★ 例文：To address these issues, we proposed a hierarchical dense positive supervision method based on RT-DETR, named RT-DETRv3.（Abstract）

superior /suːˈpɪriər/

【形容詞】優れた、上位の ★★★ 例文：Its accuracy and speed are superior to the YOLO series of the same period, and it has received widespread attention.（Related Work）

insufficient /ˌɪnsəˈfɪʃənt/

【形容詞】不十分な ★★★ 例文：However, compared to dense supervision detectors like the YOLO series, the Hungarian matching provides much sparser supervision, leading to insufficient model training and difficult to achieve optimal results.（Abstract）

additional /əˈdɪʃənəl/

【形容詞】追加の ★★ 例文：These approaches significantly improve the model's performance and accelerate convergence without additional inference latency.（Introduction）

parallel /ˈpærəlel/

【形容詞】並列の、並行した ★★ 例文：Co-DETR [33] proposed multiple parallel one-to-many label assignment auxiliary head training strategies (e.g., ATSS [30] and Faster RCNN [20]), which can easily enhance the learning ability of the encoder in end-to-end detectors.（Related Work）

relevant /ˈreləvənt/

【形容詞】関連する ★★ 例文：Additionally, to ensure that multiple relevant queries focus on the same target, we introduce a supplementary one-to-many matching branch.（Related Work）

supplementary /ˌsʌplɪˈmentəri/

【形容詞】補助的な、補足の ★★ 例文：Additionally, to ensure that multiple relevant queries focus on the same target, we introduce a supplementary one-to-many matching branch.（Related Work）

sufficient /səˈfɪʃənt/

【形容詞】十分な ★★★ 例文：This strategy can effectively strengthen the supervision of the encoder, enabling it to have sufficient representation ability to accelerate the convergence of the model.（Method）

unique /juːˈniːk/

【形容詞】唯一の、固有の ★★ 例文：Specifically, we employ a query selection module to generate a unique set of object queries.（Method）

comparable /ˈkɑːmpərəbəl/

【形容詞】匹敵する、同等の ★★ 例文：It clearly demonstrates that our method outperforms them in terms of convergence speed in any schedule and only needs half of the training epochs to achieve the comparable performance.（Experiments）

varying /ˈveriɪŋ/ （基本形：vary）

【形容詞】さまざまな、変化する ★★ 例文：To verify the effect varying the number of self-attention perturbation branches on RT-DETRv3 performance, we conducted ablation experiments using RT-DETRv3-R18 with branch counts of 2, 3, and 4, while keeping all other configurations unchanged.（Experiments）

widespread /ˈwaɪdspred/

【形容詞】広く行き渡った、広範囲の ★★ 例文：Its accuracy and speed are superior to the YOLO series of the same period, and it has received widespread attention.（Related Work）

comparative /kəmˈpærətɪv/

【形容詞】比較の ★ 例文：Table 1 presents a comparative analysis of RT-DETRv3, RT-DETR [32], and RT-DETRv2 [19] across various training schedules.（Experiments）

overall /ˈoʊvərɔːl/

【形容詞】全体の、総合的な ★★ 例文：These modules enhance the convergence speed and improve the overall performance of the RT-DETR [32].（Related Work）

responsible /rɪˈspɑːnsəbəl/

【形容詞】〜を担う、責任を持つ ★★ 例文：In summary, the overall loss function of our proposed approach is as follows: L = αLaux + βLo2o + γLo2m, where Laux is responsible for dense supervision of the encoder, Lo2o enriches the one-to-one supervision information for the decoder while preserving the end-to-end prediction characteristics, and Lo2m provides one-to-many dense supervision to the decoder.（Method）

core /kɔːr/

【形容詞】中核の、核となる ★★ 例文：We preserve the core architecture of RT-DETR (highlighted in yellow) and propose a novel hierarchical decoupled dense supervision method (emphasized in green).（Method）

default /dɪˈfɔːlt/

【形容詞】既定の、デフォルトの ★★ 例文：During the sample matching phase, an augmented target set is generated by replicating the training labels by a factor of m, with a default value of 4.（Method）

aforementioned /əˈfɔːrmenʃənd/

【形容詞】前述の ★★★ 例文：Notably, all aforementioned modules are training-only.（Abstract）

enormous /ɪˈnɔːrməs/

【形容詞】膨大な、非常に大きな ★★ 例文：It has enormous value in practical applications such as autonomous driving, video surveillance, and object tracking.（Introduction）

main /meɪn/

【形容詞】主要な、主な ★★ 例文：We reused the RT-DETR [32] decoder structure as the main branch and additionally added three groups of parameter-shared self-attention perturbation branches.（Method）

augmented /ɔːɡˈmentɪd/ （基本形：augment）

【形容詞】拡張された ★★ 例文：During the sample matching phase, an augmented target set is generated by replicating the training labels by a factor of m, with a default value of 4.（Method）

perturbed /pərˈtɝːbd/ （基本形：perturb）

【形容詞】摂動を受けた ★★★ 例文：Then, Qi and Ki are multiplied to compute the attention weight, which is further multiplied by Mi and passed through a softmax function to yield the perturbed attention weight.（Method）

additionally /əˈdɪʃənəli/

【副詞】さらに、加えて ★★★ 例文：Additionally, we introduce a shared-weight decoder branch for dense positive supervision to ensure more high-quality queries matching each ground truth.（Abstract）

firstly /ˈfɝːstli/

【副詞】第一に ★★ 例文：Firstly, we introduce a CNN-based auxiliary branch that provides dense supervision that collaborates with the original decoder to enhance the encoder's feature representation.（Abstract）

secondly /ˈsekəndli/

【副詞】第二に ★★ 例文：Secondly, to address insufficient decoder training, we propose a novel learning strategy involving self-attention perturbation.（Abstract）

notably /ˈnoʊtəbli/

【副詞】特に、とりわけ ★★★ 例文：Notably, all aforementioned modules are training-only.（Abstract）

mainly /ˈmeɪnli/

【副詞】主に ★★★ 例文：DAC-DETR [9], MS-DETR [31], and GroupDETR [4] mainly accelerate the convergence of the model by adding one-to-many supervised information to the decoder of the model.（Related Work）

thoroughly /ˈθɝːoʊli/

【副詞】徹底的に、十分に ★★ 例文：Extensive experiments conducted on the COCO dataset have thoroughly validated the effectiveness of our proposed approach.（Introduction）

easily /ˈiːzɪli/

【副詞】容易に ★★ 例文：Co-DETR [33] proposed multiple parallel one-to-many label assignment auxiliary head training strategies (e.g., ATSS [30] and Faster RCNN [20]), which can easily enhance the learning ability of the encoder in end-to-end detectors.（Related Work）

ultimately /ˈʌltɪmətli/

【副詞】最終的に、結局 ★★★ 例文：However, the adoption of a one-to-one set matching in the RT-DETR leads to sparse supervision information, ultimately impairing the model's performance.（Method）

strictly /ˈstrɪktli/

【副詞】厳密に ★★ 例文：The AdamW optimizer, integrated with a weight decay factor of 0.0001, was employed, ensuring that all other training configurations adhered strictly to the RT-DETR [32], encompassing both data augmentation and pre-training.（Method）

clearly /ˈklɪrli/

【副詞】明らかに ★★ 例文：It clearly demonstrates that our method outperforms them in terms of convergence speed in any schedule and only needs half of the training epochs to achieve the comparable performance.（Experiments）

collaboratively /kəˈlæbəreɪtɪvli/

【副詞】協調して ★ 例文：For the CNN-based one-to-many auxiliary branch, we directly employ existing state-of-the-art dense supervision methods, such as PP-YOLOE [28], to collaboratively supervise the encoder's representation learning.（Method）

conversely /kənˈvɝːsli/

【副詞】逆に ★★ 例文：Conversely, increasing the number of branches excessively raised the model's learning difficulty without yielding significant performance gains.（Experiments）

excessively /ɪkˈsesɪvli/

【副詞】過度に ★★ 例文：Conversely, increasing the number of branches excessively raised the model's learning difficulty without yielding significant performance gains.（Experiments）

matching /ˈmætʃɪŋ/ （基本形：match）

【動詞】マッチング、照合 ★★★ 例文：However, the adoption of a one-to-one set matching in the RT-DETR leads to sparse supervision information, ultimately impairing the model's performance.（Method）

positive /ˈpɑːzətɪv/

【形容詞】ポジティブな（positive sample＝正例） ★★★ 例文：This strategy diversifies label assignment for positive samples across multiple query groups, thereby enriching positive supervisions.（Abstract）

subsequently /ˈsʌbsɪkwəntli/

【副詞】その後、続いて ★★ 例文：Subsequently, each query updates itself by retrieving information from the encoder's output features via the cross-attention module.（Method）

sample /ˈsæmpəl/

【名詞】サンプル、標本（positive sample＝正例） ★★ 例文：The introduction of multiple sets of random perturbations diversifies the features of the queries, allowing multiple related queries associated with the same target to have a chance of being assigned as positive sample queries, thereby enriching the supervision information.（Method）