Hi! Thanks for making your code open. Could you please further explain why the parameter chooses 0.3? Why not dp_1 = self.sigmoid(dp_1)?

https://github.com/visinf/multi-mono-sf/blob/8e71381635c7ad4981055301705e018a884a6085/models/model_monosceneflow.py#L122

I have noticed when evaluating the pre-trained model on a dataset that has camera motion, that the model predicts large scene flow for the background objects that are actually static. As far as I can tell, this is an unintended consequence - is that correct? In theory at least this yields invalid results, since to have non 0 scene flow vectors in world space would mean that the static objects are moving.

I discovered this on my custom dataset, but it can also be seen on the Davis camel scene, in which the camel is moving but the background is static (in world coordinates). Example input frame:

An example scene flow (visualisation) output is as follows: (The red and blue squares have been manually annotated by me for the following reason)

Inspecting the values at the red (corresponding to static objects) and blue (corresponding to dynamic objects) squares, we see that the scene flow is larger at the red than at the blue:

• red region: max 0.16379195 , min: 0.024987139
• blue region: max 0.053224254 , min: 5.828061e-07

(max and min are across all three sf dimensions)

My use of this model is contingent on it's ability to predict the sf of static objects as 0. Is this a known issue, or am I missing something.

Thanks so much!

Hi Junhwa, thanks a lot for the awesome work in the field of scene flow :) When l tried to use your loss in my environment setting (cudatoolkit=11.2), the pts_loss(s_3) became extremely big, for example, bigger than 100000. But when I set the version of cudatoolkit as 10.2, the output of pts_loss became normal (usually less than 10). I can not find out the reason. Have you ever met this kind of situation? Again, thanks a lot.

Hi Junhwa When checking into a class, Eval_SceneFlow_KITTI_Train_Multi, in your code, I found two weird constants: 91875.68 and 91873.4. It does not make sense to me. Why not use something like "gt_*_mask.sum()" here?

Your research is so interesting. (selfmono-sf and multimono-sf) Your depth metric does not computing median scaling. But, another KITTI depth estimation model, monodepth2 uses median scaling computation

Why is there a difference in the metric computation method?

Your code multimono-sf https://github.com/visinf/self-mono-sf/blob/054df08674a2df96885682e657ca4803c129a364/losses.py#L402 selfmono-sf https://github.com/visinf/multi-mono-sf/blob/main/losses.py#L513

Monodepth2 KITTI depth metric https://github.com/nianticlabs/monodepth2/blob/master/evaluate_depth.py#L206

I cloned the code and even installed it.

When I run "bash train_selfsup.sh" in scripts, the following error appears: "unrecognized arguments: where checkpoints and log files will be saved)/MonoSceneFlow_Multi-selfsup-20210722-135602 Raw dataset path) Raw dataset path)" Is there any additional folder configuration I need to set up?

From the figure 1. (in your paper):

I does not get where you use the stereo images. and your evaluation code does not use them (the davis dataset...)

And can I do self-supervised training without stereo images? Since many wild datasets (open world dataset) do not have them.

Hello, after I modified the dataset address ，run ./eval_ft_test.sh , report an error： mian.py:error : unrecognized arguments: --validation_dataset_preprocessing_crop=False. Why is this?Hope to get your answer.