about OSCD

[Larry-Zheng]

Could you tell me your best metrics on OSCD please? Thank you! The official training code get best f1 score at 60.15%. I designed a network with less parameters getting slightly better f1 score at 61%. But I've read a paper using just plain unet which reports f1 score at 75%. I'm a little suspicious and confused. So I just want to ask your opinion on this.

[Larry-Zheng]

兄嘚 刚刚看了你github 你应该是国人 我还是说明白一点吧 我是这几天一直在跑OSCD 官方的那个代码最多就train到60% 然后看到一篇文章 叫啥train at cern 貌似是毕业论文的样子 说train到75%了 很迷 所以想问问你 你用你的代码跑OSCD最多跑到多少了

[Larry-Zheng]

方便 加个Q吗 我Q：519334736

[Bobholamovic]

说实话，我刚刚翻了下实验记录，我好像没在OSCD数据集上做过实验……那时候把接口写好之后就没动过了😂

[Larry-Zheng]

说实话，我刚刚翻了下实验记录，我好像没在OSCD数据集上做过实验……那时候把接口写好之后就没动过了😂

哦哦好的 还是想问一下 另外几个数据集你试验出来效果如何

[Bobholamovic]

在Air Change数据集上按照原作者的结构我并没有取得很好的效果（也可能是超参数没调好），不过把通道数加倍，dropout去掉之后，得到的结果大概是这样：

Szada

Arch	Precision	Recall	F1-score	OA
FC-EF	0.46	0.64	0.53	0.93
FC-Siam-conc	0.42	0.58	0.49	0.93

Tiszadob

Arch	Precision	Recall	F1-score	OA
FC-EF	0.65	0.68	0.67	0.92
FC-Siam-conc	0.63	0.66	0.65	0.91

Air Change数据集太小，所以对超参数过于敏感，随便动一下对最后指标影响可能都蛮大。

Season-varying 数据集（这个数据集不推荐使用，原因见这里，另我怀疑这个数据集训练集和测试集有重叠所以指标虚高）:

Arch	Precision	Recall	F1-score	OA
FC-EF	0.83	0.98	0.90	0.98
FC-Siam-conc	0.85	0.99	0.91	0.98
FC-Siam-diff	0.87	0.98	0.92	0.98

还可以参考其它论文里的指标：

[1] Peng, D., Zhang, Y., & Guan, H. (2019). End-to-End Change Detection for High Resolution Satellite Images Using Improved UNet++. Remote Sensing, 11(11), 1382.
[2] Chen, H., & Shi, Z. (2020). A Spatial-Temporal Attention-Based Method and a New Dataset for Remote Sensing Image Change Detection. Remote Sensing, 12(10), 1662.
[3] H. Chen, C. Wu, B. Du, & L. Zhang (2019). Deep Siamese Multi-scale Convolutional Network for Change Detection in Multi-temporal VHR Images. In 2019 10th International Workshop on the Analysis of Multitemporal Remote Sensing Images (MultiTemp) (pp. 1–4).
[4] Chen, J., Yuan, Z., Peng, J., Chen, L., Huang, H., Zhu, J., et al. (2020). DASNet: Dual attentive fully convolutional siamese networks for change detection of high resolution satellite images.

[Larry-Zheng]

好的 谢谢啦 兄嘚~

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人: "Lin Manhui"<[email protected]&gt;; 发送时间: 2020年7月8日(星期三) 上午10:49 收件人: "Bobholamovic/FCN-CD-PyTorch"<[email protected]&gt;; 抄送: "bittersweet"<[email protected]&gt;; "Author"<[email protected]&gt;; 主题: Re: [Bobholamovic/FCN-CD-PyTorch] about OSCD (#4) 在Air Change数据集上按照原作者的结构我并没有取得很好的效果（也可能是超参数没调好），不过把通道数加倍，dropout去掉之后，得到的结果大概是这样： Szada Arch Precision Recall F1-score OA FC-EF 0.46 0.64 0.53 0.93 FC-Siam-conc 0.42 0.58 0.49 0.93 Tiszadob Arch Precision Recall F1-score OA FC-EF 0.65 0.68 0.67 0.92 FC-Siam-conc 0.63 0.66 0.65 0.91 Air Change数据集太小，所以对超参数过于敏感，随便动一下对最后指标影响可能都蛮大。 Season-varying 数据集（这个数据集不推荐使用，原因见这里，另我怀疑这个数据集训练集和测试集有重叠所以指标虚高）: Arch Precision Recall F1-score OA FC-EF 0.83 0.98 0.90 0.98 FC-Siam-conc 0.85 0.99 0.91 0.98 FC-Siam-diff 0.87 0.98 0.92 0.98 还可以参考其它论文里的指标： [1] Peng, D., Zhang, Y., &amp; Guan, H. (2019). End-to-End Change Detection for High Resolution Satellite Images Using Improved UNet++. Remote Sensing, 11(11), 1382. [2] Chen, H., &amp; Shi, Z. (2020). A Spatial-Temporal Attention-Based Method and a New Dataset for Remote Sensing Image Change Detection. Remote Sensing, 12(10), 1662. [3] H. Chen, C. Wu, B. Du, &amp; L. Zhang (2019). Deep Siamese Multi-scale Convolutional Network for Change Detection in Multi-temporal VHR Images. In 2019 10th International Workshop on the Analysis of Multitemporal Remote Sensing Images (MultiTemp) (pp.&nbsp;1–4). [4] Chen, J., Yuan, Z., Peng, J., Chen, L., Huang, H., Zhu, J., et al. (2020). DASNet: Dual attentive fully convolutional siamese networks for change detection of high resolution satellite images. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.