Questions about optical flow groundtruth

[Hazeliii]

Hi Alex, thanks for your work.

I'd like to know the measurement of the optical flow data. It's hard to express my question clearly due to my poor English, but I will try my best.

In the DSEC dataset, it is said that

Optical flow data is provided at 10 Hz as displacement fields between two timestamps that are 100 milliseconds apart

If the optical flow data in (x, y) is given as (-12.67, -10.45), I would know that it means the displacement happened in 100 milliseconds.

But I did't find the explanation in the MVSEC dataset. So given the optical flow data in MVSEC , what's the time interval for the displacement?

I hope that I made the problem clearly.

Best wishes.

[alexzzhu]

Our dataset is set up similarly, and computes pixel displacements over a time interval. However, the length of the time interval is specified at ground truth generation time, so it depends on the parameters you use. For more information see our repo here: https://github.com/daniilidis-group/mvsec/tree/master/tools/gt_flow

[Hazeliii]

Thanks for you reply.

I didn't generate the optical flow data myself. I downloaded the ground truth from the link provided by MVSEC dataset. So the time interval is specified , right? Would you please tell me the time interval?

And I checked the code in https://github.com/daniilidis-group/mvsec/tree/master/tools/gt_flow . In the extract_single_bag.py , as I can see, the parameters would't affect the time interval.

According to the paper and the code, my understanding is that the time interval depends on the delta_T between two camera poses.
The paper says，

From the pose (consisting of rotation R and translation p) of the camera at time t0 and t1 ....

And the code is as follows,

p0, q0, t0 = self.p_q_t_from_msg(P0)
p1, q1, t1 = self.p_q_t_from_msg(P1)
...
dt = t1 - t0

......
flat_x_flow_out *= dt * self.P[0,0]
flat_y_flow_out *= dt * self.P[1,1]

I think the delta_T between two camera poses decides the time interval of optical flow, and the delta_T between two camera poses is the same for all the pose ground truth.

Those above are just my understanding. I hope you can tell me the correct answers.

Best wishes.

[alexzzhu]

Ah I see. In that case, the delta_T in this case will be the time difference between successive timestamps in the npz file. I think an approximation of 10Hz would be close enough here.

[Hazeliii]

From the optical flow npz file, the time difference between successive timestamps is 50ms instead of 100ms（10HZ）. I hope that my understanding is correct.

1.506120259672562599e+9
1.506120259722571135e+9
1.506120259772649050e+9
1.506120259822633266e+9
1.506120259872704029e+9
1.506120259922727823e+9
1.506120259972785950e+9
1.506120260022795200e+9
1.506120260072817564e+9
1.506120260122830391e+9
1.506120260173165560e+9
1.506120260222922564e+9

I am sorry for keeping asking this, but it is important for me . I want to make sure.
Thanks a lot for you patience.