seokjulee / insta-dm Goto Github PK

Dear authors

I was wondering how you obtain the backward flow of the first frame and the forward flow of the last frame in a scene.
Ex: there are 000.jpg, 001.jpg, 002.jpg in a scene
000.jpg, 001.jpg => forward_000.flo
(no -001.jpg to compute backward_000.flo)
001.jpg, 002.jpg => forward_001.flo
001.jpg, 000.jpg => backward_001.flo
002.jpg, 001.jpg => backward_002.flo
(no 003.jpg to compute backward_002.flo)

Thank you very much.

Best,
Bolian

Hi, thank you for sharing this amazing work, I just have a little bit confusion. Does the AbsRel error of 0.112 on kitti eigen split with or without imagenet pretraining? If not, do you have results with pretraining?

Thank you

How are the pre-training weights of each sub-network obtained? How are the sub-networks supervised learning?

I've read the paper and it's amazing work when would plan to make the repo publicly available? Thank you!

Is it possible to use this method in realtime on a video stram from a webcam?

By running sh scripts/run_demo.sh, I get the following error:

/torch/lib/python3.6/site-packages/torch_sparse/_version_cuda.so: undefined symbol: ZN3c106detail12infer_schema20make_function_schemaENS_8ArrayRefINS1_11ArgumentDefEEES4

i install exactly
pytorch=1.7.0 torchvision torchaudio cudatoolkit=11.0 -c pytorch
and
pip3 install torch-scatter torch-sparse -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html

Hi, Seokju!
I get only 1636 iterms, but the "train.txt" and "val.txt" indicate there should be 4883+602 = 5485.
Did I make a mistake?

I found that the motion of objects in the code is 3-Dof, but the paper claims that it estimates 6-Dof motion of objects. So where is the 6-Dof motion estimation of objects? The prediction of ego pose is 6-Dof, but the prediction of object pose is 3-Dof

Hi @SeokjuLee :

Thank you for sharing this amazing work, the visualization in this code helps me a lot in understanding the logic. However, I have a confusion, the line below:

There are two instance masks in r2t_inst, are you supposed to visualize one of them like r2t_inst[:, :, 0] or do you have other implementations to visualize both of them in once?

Thank you!

Hi @SeokjuLee

I noticed that the Cityscape-VIS dataset does not include the GT depth, could you share a link to download the GT data as used in your paper? Thank you!

Dear authors:
Thank you for your great work in this paper ! If I want to reproduce it on my custom dataset, how should I preprocess my image data to obtain the instance segmentation results.