search

NVlabs / few-shot-vid2vid

Pytorch implementation for few-shot photorealistic video-to-video translation.
Other
1.79k stars 276 forks source link

How to pre-process pose dataset #58

Closed akkinenic closed 3 years ago

akkinenic commented 4 years ago

Can someone share how to pre-process pose dataset?

hungsing92 commented 4 years ago

Install openpose and densepose，and run preprocess.py

Jarvisss commented 3 years ago

I only use Openpose and do not configure Densepose, and here's my pre-process solution on Ubuntu 18.04:

Install Openpose

  • follow the official instruction from here

Run preprocess.py

python preprocess.py --steps openpose,clean,divide_sequences --video_root /your/path/to/videos/ --output_root /your/path/to/output --openpose_root /your/path/to/openpose

--openpose_root is where your openpose is installed, under which you will find 'build/example/opepose/openpose.bin'

update

I ran with --steps openpose, clean, divide_sequence, but all the openpose results were deleted, as the code https://github.com/NVlabs/few-shot-vid2vid/blob/bc1135071c9a3b166291dfab8cc3314a50585f09/data/preprocess/preprocess.py#L64

check for densepose result, if there are no densepose result in the folder, it will delete the frame.

So some part of my pre-processed data were gone..

I suggest the one who wants to pre-process to install both the Densepose and Openpose, otherwise the code should be modified before you do the clean step

guofengming11 commented 3 years ago

I only use Openpose and do not configure Densepose, and here's my pre-process solution on Ubuntu 18.04:

Install Openpose

  • follow the official instruction from here

Run preprocess.py

python preprocess.py --steps openpose,clean,divide_sequences --video_root /your/path/to/videos/ --output_root /your/path/to/output --openpose_root /your/path/to/openpose --openpose_root is where your openpose is installed, under which you will find 'build/example/opepose/openpose.bin'

update

I ran with --steps openpose, clean, divide_sequence, but all the openpose results were deleted, as the code https://github.com/NVlabs/few-shot-vid2vid/blob/bc1135071c9a3b166291dfab8cc3314a50585f09/data/preprocess/preprocess.py#L64

check for densepose result, if there are no densepose result in the folder, it will delete the frame.

So some part of my pre-processed data were gone..

I suggest the one who wants to pre-process to install both the Densepose and Openpose, otherwise the code should be modified before you do the clean step

Could I only use openpose ，then start training，how to start like that？