zzwei1
commented
3 years ago Hello, thank you for such valuable work. I read and debugged your code, and I have a question about how to visualize the optical flow calculated by the motionsqueeze module. I observed that the data range of the output optical flow is [-1, 1], which does not match the value range of optical flow in the traditional sense. After I converted tensor to numpy, I used the conversion function here https://github.com/princeton-vl/RAFT/blob/master/core/utils/flow_viz.py. But the effect of direct visualization does not seem meaningful. So I want to ask you how to correctly convert the optical flow into a three-channel color picture.
Hi, have you figure out how to visualize the learned flow ? I have the same question.
Is it related to the normalization of the learned flow, as in the function "match_to_flow_soft"?
In the code, the learned flow_x and flow_y are normalized with (self.patch_dilation * displacement).
rayush7
Hello, thank you for such valuable work. I read and debugged your code, and I have a question about how to visualize the optical flow calculated by the motionsqueeze module. I observed that the data range of the output optical flow is [-1, 1], which does not match the value range of optical flow in the traditional sense. After I converted tensor to numpy, I used the conversion function here https://github.com/princeton-vl/RAFT/blob/master/core/utils/flow_viz.py. But the effect of direct visualization does not seem meaningful. So I want to ask you how to correctly convert the optical flow into a three-channel color picture.