search

facebookresearch / DensePose

A real-time approach for mapping all human pixels of 2D RGB images to a 3D surface-based model of the body
http://densepose.org
Other
6.95k stars 1.29k forks source link

Converting to Tensorflow #135

Open TimoK93 opened 5 years ago

TimoK93 commented 5 years ago

Hey folks,

is there an availability to convert your DensePose model to a Tensorflow model? I'd like to work on your results for my Master Thesis, but my framework is based on Tensorflow...

I'm pretty new in Caffe2 and so one more question: Is it possible to run the model as standalone without installing Detectron/Densepose framework?

roboticlemon commented 5 years ago

It is absolutely not possible. Everything is done through Caffe2. Whilst it's possible you could create a network in TF with the dataset, annotations and general architecture, it wouldn't be worth your time!

TimoK93 commented 5 years ago

Okay, sounds sensible!

Is it possible to run the net as a "standalone" in Caffe2 without Densepose/Detectron? I'm thinking about something like (heavy simplified):

import caffe2 image = load_image(test_image_path) net = load_net(densepose_net_path result = net.predict(image) do_something_with_the_result()

I want to do research on the net and therefore i need to do inference in automated frameworks. Is the only possibility to do inference with the command line tool described in GettingStarted.md?

roboticlemon commented 5 years ago

The "command line tool" is just a Python wrapper.

If u look through infer_simple and train_net you will find the inference and training. If you mean testing as in testing against ground truth annotations you can have a look through test_net

I'm not sure what you mean about without DensePose/Detectron. Detectron is the vision framework within caffe2, it handles just about everything vision based. DensePose as is in this repo has a series of modifications to Detectron that allow for visualising, training and testing for body_UV coordinates (dense pose estimation)

You can absolutely do what you want to do though!

RoboEvangelist commented 5 years ago

https://github.com/llSourcell/pose_estimation