| Item type |
SIG Technical Reports(1) |
| 公開日 |
2023-02-23 |
| タイトル |
|
|
タイトル |
Dynamic Weight Scheduling for Long-tailed Visual Recognition |
| タイトル |
|
|
言語 |
en |
|
タイトル |
Dynamic Weight Scheduling for Long-tailed Visual Recognition |
| 言語 |
|
|
言語 |
eng |
| 資源タイプ |
|
|
資源タイプ識別子 |
http://purl.org/coar/resource_type/c_18gh |
|
資源タイプ |
technical report |
| 著者所属 |
|
|
|
Ritsumeikan University |
| 著者所属 |
|
|
|
Hitotsubashi University |
| 著者所属 |
|
|
|
Ritsumeikan University |
| 著者所属(英) |
|
|
|
en |
|
|
Ritsumeikan University |
| 著者所属(英) |
|
|
|
en |
|
|
Hitotsubashi University |
| 著者所属(英) |
|
|
|
en |
|
|
Ritsumeikan University |
| 著者名 |
Xinyuan, Li
Yu, Wang
Jien, Kato
|
| 著者名(英) |
Xinyuan, Li
Yu, Wang
Jien, Kato
|
| 論文抄録 |
|
|
内容記述タイプ |
Other |
|
内容記述 |
To address the issue of static re-weighting methods negatively impacting the accuracy of head categories in long-tailed image recognition tasks, this paper introduces a dynamic re-weighting method that adjusts the weight assigned to a specific category based on the number of samples that have been used for that category. This weighting scheme increases and then decreases in proportion to the sample usage. Additionally, the use of our head-to-tail loss helps control the evolution of weights, enabling the model to progressively shift its focus from head categories to tail categories. Experiments on long-tailed CIFAR/ImageNet datasets demonstrate that this approach outperforms static re-weighting methods and improves the accuracy of tail categories without diminishing the accuracy of head categories. In addition, we introduce our dynamic re-weighting method in two-stage training and discuss the advantages and limitations of our method. |
| 論文抄録(英) |
|
|
内容記述タイプ |
Other |
|
内容記述 |
To address the issue of static re-weighting methods negatively impacting the accuracy of head categories in long-tailed image recognition tasks, this paper introduces a dynamic re-weighting method that adjusts the weight assigned to a specific category based on the number of samples that have been used for that category. This weighting scheme increases and then decreases in proportion to the sample usage. Additionally, the use of our head-to-tail loss helps control the evolution of weights, enabling the model to progressively shift its focus from head categories to tail categories. Experiments on long-tailed CIFAR/ImageNet datasets demonstrate that this approach outperforms static re-weighting methods and improves the accuracy of tail categories without diminishing the accuracy of head categories. In addition, we introduce our dynamic re-weighting method in two-stage training and discuss the advantages and limitations of our method. |
| 書誌レコードID |
|
|
収録物識別子タイプ |
NCID |
|
収録物識別子 |
AA11131797 |
| 書誌情報 |
研究報告コンピュータビジョンとイメージメディア（CVIM）
巻 2023-CVIM-233,
号 25,
p. 1-4,
発行日 2023-02-23
|
| ISSN |
|
|
収録物識別子タイプ |
ISSN |
|
収録物識別子 |
2188-8701 |
| Notice |
|
|
|
SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc. |
| 出版者 |
|
|
言語 |
ja |
|
出版者 |
情報処理学会 |
IPSJ-CVIM23233025.pdf (1.4 MB)
