some question about code

[xiaobai65]

Thanks for your great job！I am currently researching weakly supervised VSOD. As mentioned in your article, you conducted testing on the VSOD task. Could you please provide the code for the preprocessing section, (Code for Expanding Scribble Annotations Using the SILC Algorithm)，and give me some training detail in this task? I look forward to your reply！thanks！

[liuzywen]

Code for Expanding Scribble Annotations Using the SILC Algorithm can be seen data.py.

[liuzywen]

训练策略与"Weakly Supervised Video Salient Object Detection"论文保持一致。
训练细节：
训练分为两个阶段，预训练和微调阶段。
预训练阶段利用显著数据集S-DUTS，利用现有的光流估计方法"Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume."获得image对应的光流图。
微调阶段利用DAVIS和DAVSOD训练集，训练模型。
图像分辨率是256*256，优化器使用的AdamW，学习率1e-5，batchsize设置为24，数据增强采用随机翻转和裁剪，预训练和微调epoch分别是40和60。

[xiaobai65]

多谢！请问您现在还保留有生成的光流图吗？因为我现在复现不出wsvsod的结果，我怀疑是不是我生成光流的代码有些许问题，如果您还保留有生成的光流图，能以Goggle drive或者网盘的形式放到您的github上吗？

…

---- 回复的原邮件 ---- | 发件人 | ***@***.***> | | 日期 | 2023年05月12日 07:56 | | 收件人 | ***@***.***> | | 抄送至 | ***@***.***>***@***.***> | | 主题 | Re: [liuzywen/RGBTScribble-ICME2023] some question about code (Issue #1) | 训练策略与"Weakly Supervised Video Salient Object Detection"论文保持一致。 训练细节： 训练分为两个阶段，预训练和微调阶段。 预训练阶段利用显著数据集S-DUTS，利用现有的光流估计方法"Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume."获得image对应的光流图。 微调阶段利用DAVIS和DAVSOD训练集，训练模型。 图像分辨率是256*256，优化器使用的AdamW，学习率1e-5，batchsize设置为24，数据增强采用随机翻转和裁剪，预训练和微调epoch分别是40和60。 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[liuzywen]

Please see https://github.com/sniklaus/pytorch-pwc to generate the optical flow map.

[xiaobai65]

ok，thanks for your reply!

…

---- Replied Message ---- | From | ***@***.***> | | Date | 05/12/2023 09:51 | | To | ***@***.***> | | Cc | ***@***.***>***@***.***> | | Subject | Re: [liuzywen/RGBTScribble-ICME2023] some question about code (Issue #1) | Please see https://github.com/sniklaus/pytorch-pwc to generate the optical flow map. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>