Static and dynamic decomposing quality

[ntaquan0125]

Dear authors,

I attempted to re-conduct the experiments on the Virtual KITTI dataset and observed that the quality of decomposition is good only within the ego-static sequence. However, when the ego car is in motion, it degrades significantly, as illustrated below:

*The model was trained on Scene06 using the file configs/vkitti2-06.yaml

Fig. 1: Clear separation (frame 0)

Fig. 2: Bad separation (frame 267)

Fig. 3: Forward flow (frame 267)

In my opinion, the flow estimated from ego motion is mainly the cause here.
Could you tell me what adjustment I need to do in order to achieve better separation in this case?

The modification I made in the configuration to suite my hardware:

• train_downscale_factor = 2
• train_num_rays_per_batch = 3000
• use correspondences to calculate flow

Thank you!