BO-LIAN CHEN

@wangfrombupt Thanks for mentioning this bug which I did not noticed before. I guess the original codes assumes the feature map (x) always has even width and height. I would...

Dear authors I just noticed the number of images in a scene is one less than the numbers of either forward or backward flows. With the same example (000.jpg, 001.jpg,...

Dear @feipan664 Thanks very much for pointing out this problem. We have been noticed it since the official TensorFlow repository. It is an unsolved issue, so I would like to...

The "blackholes" issue is an intrinsic problem for unsupervised depth estimation methods whose main supervision is based on image reconstruction. While objects having the same velocity with the camera, they...

Dear @feipan664 It would probably take longer to tweak losses and do experiments to check their effects. I would like to share the workaround first. We found the "blackholes" issue...

Hi @wanglong1008 Thanks for the question. Yes, the pose_net is supposed to output relative pose of two images. Would you please share how you evaluate the pose? (sample code or...

Hi @wanglong1008 Sincerely apologize for my late response. I marked this issue as help wanted as a future check item.

Hi @18811598673 感謝您的提問，您提出的不等式是作者在Appendix A.3，基於一些假設所推導的fx與fy的誤差(δfx, δfy)上界。 該章節主要想探討隨著R接近I (i.e.沒有旋轉)，約束變弱的狀況下，對fx與fy的誤差有何影響： 左邊的不等式代表只有y軸旋轉(ry)時的fx誤差; 右式則代表只有x軸旋轉時的fy誤差。兩者都與旋轉的倒數呈正比，證實旋轉越小誤差的上界會變大。(但我不太明白作者推導最後一步，假設(A8)右邊

是的，更準確地來說，這式子是關聯前後兩frames的影像座標(p, p')，損失是來自於兩兩關聯座標的像素差值 Σ|Image(p) - Image(p')|

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...