Different flows for the same data with multiple runs

[rfww]

Hi, Shiba. Thanks for your work. There is an issue that the estimated flow of the same event data is different under multiple runs. Why? I computed the difference between two flows and it's very big. The color wheel is the same one. How can I address it? Thanks.

[shiba24]

Hi @rfww, I do not think that is expected. From my experience this method always coverges and the difference among multiple runs is very small. Could you please provide a minimal example to reproduce it?

[rfww]

Thanks for your quick response. Some cases are as follows:

These are estimated from the same event data (from dsec dataset) with three runs.

[rfww]

For efficiency, I sampled 30000 events for the above case. The following cases are tested on the whole event stream:

The event data can be downloaded at [event.npy](https://github.com/rfww/files/blob/main/event.npy). Hope you can tell me where I went wrong.

[shiba24]

Can you also provide your config.yaml file?

[rfww]

Config file is still the mvsec_indoor_no_timeaware.yaml. To input dsec data, I just rewrite lines 171-176 in main.py as:
batch= np.load("event.npy") # x y t p
batch= np.array(batch).astype(np.float32)
last_stamp = batch[-1, 2]
first_stamp = batch[0, 2]
batch[:, 2] = (batch[:, 2]-first_stamp)/(last_stamp-first_stamp)
solv.visualize_one_batch_warp(batch)
best_motion: np.ndarray = solv.optimize(batch)
solv.visualize_one_batch_warp(batch, best_motion)

[rfww]

Hi Shiba, have any suggestions for this issue? Thanks!

[shiba24]

Hi @rfww I do not have time to check details but if you are using the same config as MVSEC that might be the cause. You can adjust image height and width and cropped ones, as well.

[shiba24]

@rfww did you solve the issue? Please provide your config file