Question

[KOuldamer]

Thank you for this interesting work,

I want to use this method to estimate camera calibration parameters. Can it work on fixed camera videos in scenes with a lot of movement?

[bolianchen]

Hi @KOuldamer I haven't tried training in scenarios as you mentioned, but I guess it is challenging according to the design.

Assume the model could accurately learn the camera pose is 0's and the relative motions of objects. Since background is static and does not inject any information to update the networks via image reconstruction loss, it is likely to degrade the whole learning process (including camera parameters) and break the initial assumption made above.

Best, Bolian

[zhangzhenfengjy]

Hi, can this just use a separate RGB image? Or does it require additional pose input?

[zhangzhenfengjy]

Thank you for this interesting work,

I want to use this method to estimate camera calibration parameters. Can it work on fixed camera videos in scenes with a lot of movement? Hi, can this just use a separate RGB image? Or does it require additional pose input?

[bolianchen]

@zhangzhenfengjy You can input a separate RGB image to obtain its depth map without pose input. Please check the Run Inference section. https://github.com/bolianchen/pytorch_depth_from_videos_in_the_wild#run-inference.

[zhangzhenfengjy]

I got it.thank you very much.

发自我的iPhone

------------------ Original ------------------ From: BO-LIAN CHEN @.> Date: Thu,Jul 14,2022 6:47 PM To: bolianchen/pytorch_depth_from_videos_in_the_wild @.> Cc: zhangzhenfengjy @.>, Mention @.> Subject: Re: [bolianchen/pytorch_depth_from_videos_in_the_wild] Question (Issue #11)

@zhangzhenfengjy You can input a separate RGB image to obtain its depth map without pose input. Please check the Run Inference section. https://github.com/bolianchen/pytorch_depth_from_videos_in_the_wild#run-inference.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

[zhangzhenfengjy]

您好，我想问一下您的数据集是怎么下载的呢？我的使用您给的命令是连接不上的。

发自我的iPhone

------------------ Original ------------------ From: BO-LIAN CHEN @.> Date: Thu,Jul 14,2022 6:47 PM To: bolianchen/pytorch_depth_from_videos_in_the_wild @.> Cc: zhangzhenfengjy @.>, Mention @.> Subject: Re: [bolianchen/pytorch_depth_from_videos_in_the_wild] Question (Issue #11)

@zhangzhenfengjy You can input a separate RGB image to obtain its depth map without pose input. Please check the Run Inference section. https://github.com/bolianchen/pytorch_depth_from_videos_in_the_wild#run-inference.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

[bolianchen]

@zhangzhenfengjy

[在專案資料夾下執行以下指令，我剛試過都還是可以下載的 ./datasets/data_prep/kitti_raw_downloader.sh

[zhangzhenfengjy]

您好，这个预测的深度信息是怎么进行约束的呢？深度的损失是什么呀？请您为我解惑，谢谢🙏

发自我的iPhone

------------------ Original ------------------ From: BO-LIAN CHEN @.> Date: Mon,Jul 18,2022 6:23 PM To: bolianchen/pytorch_depth_from_videos_in_the_wild @.> Cc: zhangzhenfengjy @.>, Mention @.> Subject: Re: [bolianchen/pytorch_depth_from_videos_in_the_wild] Question (Issue #11)

@zhangzhenfengjy

[在專案資料夾下執行以下指令，我剛試過都還是可以下載的 ./datasets/data_prep/kitti_raw_downloader.sh

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

[bolianchen]

@zhangzhenfengjy 抱歉晚回覆您了！ 建議您看SfMLearner論文，對於loss會有比較清楚的說明。 簡單來說，就是根據Structure from Motion原理，用depth network + pose network + camera intrinsics，得出前後禎畫面像素點的關聯性，把預測深度的問題轉化為前後禎畫面的重建問題。而Depth from Videos in the Wild則是緩解SfM方法要求畫面中不能有運動物體的限制，針對所有可能運動物體，額外預測3個軸向的平移，另外也提供在訓練過程中習得camera intrinsics的選擇。

[zhangzhenfengjy]

感谢您的建议，数据集我还是下载不了，您有什么好办法可以分享一下吗？我不确定是否是我的网络原因，感谢🙏

发自我的iPhone

------------------ Original ------------------ From: BO-LIAN CHEN @.> Date: Mon,Jul 25,2022 10:56 AM To: bolianchen/pytorch_depth_from_videos_in_the_wild @.> Cc: zhangzhenfengjy @.>, Mention @.> Subject: Re: [bolianchen/pytorch_depth_from_videos_in_the_wild] Question (Issue #11)

@zhangzhenfengjy 抱歉晚回覆您了！ 建議您看SfMLearner論文，對於loss會有比較清楚的說明。 簡單來說，就是根據Structure from Motion原理，用depth network + pose network + camera intrinsics，得出前後禎畫面像素點的關聯性，把預測深度的問題轉化為前後禎畫面的重建問題。而Depth from Videos in the Wild則是緩解SfM方法要求畫面中不能有運動物體的限制，針對所有可能運動物體，額外預測3個軸向的平移，另外也提供在訓練過程中習得camera intrinsics的選擇。

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

[bolianchen]

@zhangzhenfengjy 可以到官方網站註冊後，就能下載資料http://www.cvlibs.net/datasets/kitti/raw_data.php

[Ryan7180]

谢谢你的建议，我还是下载不了，你有什么好的办法可以分享一下吗？我认为数据是我的网络原因吗？ 发自己的iPhone … ------------------ Original ------------------ From: BO-LIAN CHEN @.> Date: Mon,Jul 25,2022 10:56 AM To: bolianchen/pytorch_depth_from_videos_in_the_wild @.> Cc: zhangzhenfengjy @.>, Mention @.> Subject: Re: [bolianchen/pytorch_depth_from_videos_in_the_wild] Question (Issue #11) @zhangzhenfengjy 抱歉晚回覆您了！ 建議您看SfMLearner論文，對於loss會有比較清楚的說明。 簡單來說，就是根據Structure from Motion原理，用depth network + pose network + camera intrinsics，得出前後禎畫面像素點的關聯性，把預測深度的問題轉化為前後禎畫面的重建問題。而Depth from Videos in the Wild則是緩解SfM方法要求畫面中不能有運動物體的限制，針對所有可能運動物體，額外預測3個軸向的平移，另外也提供在訓練過程中習得camera intrinsics的選擇。 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

如果你是在大陆，不翻墙的话估计是下不了，很多事都干不了，ubuntu建议下个Pigchar买个无限流量就可以了，win系统好用的梯子就比较多了