Closed chengrongliang closed 4 years ago
Hi, I used the version, 4.40.0. Oh right, this error occurred to me when I used the recent version. I recommend you to downgrade if it's okay in your local setting :)
@hurjunhwa It's ok, awesome work!!! By the way, how did you convert to disp_1 frame to disp_0 frame(reference frame) when evaluation in kitti benchmark server?
disp_0: Disparity maps of first image pair in reference frame (first left image); needed for the stereo and scene flow benchmark. disp_1: Disparity information of second image pair mapped into the reference (!) frame (first left image) via the optical flow; required only for submissions to the scene flow benchmark (for specifying the scene flow of every pixel in the reference frame, we specify the disparity in the first and second image and the optical flow, all represented in the reference frame! If your method represents the disparity estimation of the second image pair in the second left image, then you need to map it back to the first image and fill in the missing values).
You can refer from this line in losses.py
.
https://github.com/visinf/self-mono-sf/blob/5f4e07955351658fa0060e6ecadca6167693b09d/losses.py#L488
From the output depth (say, depth_0) and scene flow in the reference frame, we calculate the depth value of each pixel at the next time step, by simply adding the z-component of the estimated scene flow, (say, depth_1). Then using the baseline distance of the stereo rig, those depth maps (depth_0, depth_1) are converted to the disparity maps (disp_0, disp_1)
From the description of KITTI dataset, disp_1 is not quite intuitive, but you can think that it contains the depth value of each pixel (in the reference frame) in the next frame.
@hurjunhwa OK, I got it, thank you very much.
Which is your tqdm version, error occur!!!