PoseFromShape

(BMVC 2019) PyTorch implementation of Paper "Pose from Shape: Deep Pose Estimation for Arbitrary 3D Objects" [PDF] [Project webpage]

If our project is helpful for your research, please consider citing:

@INPROCEEDINGS{Xiao2019PoseFromShape,
    author    = {Yang Xiao and Xuchong Qiu and Pierre{-}Alain Langlois and Mathieu Aubry and Renaud Marlet},
    title     = {Pose from Shape: Deep Pose Estimation for Arbitrary {3D} Objects},
    booktitle = {British Machine Vision Conference (BMVC)},
    year      = {2019}}

Updates

Release training code with Point Cloud.

See run/train_Pascal3D.sh and run/train_ObjectNet3D.sh for updated running script.

See data/README.sh for more details in data preparation.

Installation

Dependencies

The code can be used in Linux system with the the following dependencies: Python 3.6, Pytorch 1.0.1, Python-Blender 2.77, meshlabserver

We recommend to utilize conda environment to install all dependencies and test the code.

## Download the repository
git clone 'https://github.com/YoungXIAO13/PoseFromShape'
cd PoseFromShape

## Create python env with relevant packages
conda create --name PoseFromShape --file auxiliary/spec-file.txt
source activate PoseFromShape
conda install -c conda-forge matplotlib

## Install blender as a python module
conda install auxiliary/python-blender-2.77-py36_0.tar.bz2

Datasets and Models

To download and prepare the datasets for training and testing (Pascal3D, ObjectNet3D, ShapeNetCore, SUN397, Pix3D, LineMod):

cd data
bash prepare_data.sh

To download the pretrained models (Pascal3D, ObjectNet3D, ShapeNetCore):

cd model
bash download_models.sh

Training

To train on the ObjectNet3D dataset with real images and coarse alignment:

bash run/train_ObjectNet3D.sh

To train on the Pascal3D dataset with real images and coarse alignment:

bash run/train_Pascal3D.sh

To train on the ShapeNetCore dataset with synthetic images and precise alignment:

bash run/train_ShapeNetCore.sh

Testing

While the network was trained on real or synthetic images, all the testing was done on real images.

ObjectNet3D

bash run/test_ObjectNet3D.sh

You should obtain the results in Table 1 in the paper (*indicates testing on the novel categories):

Method	Average	bed	bookcase	calculator	cellphone	computer	door	cabinet	guitar	iron	knife	microwave	pen	pot	rifle	shoe	slipper	stove	toilet	tub	wheelchair
StarMap	56	73	78	91	57	82	-	84	73	3	18	94	13	56	4	-	12	87	71	51	60
StarMap*	42	37	69	19	52	73	-	78	61	2	9	88	12	51	0	-	11	82	41	49	14
Ours(MV)	73	82	90	95	65	93	97	89	75	52	32	95	54	82	45	67	46	95	82	67	66
Ours(MV)*	62	65	90	88	65	84	93	84	67	2	29	94	47	79	15	54	32	89	61	68	39

Pascal3D+

To test on the Pascal3D dataset with real images:

bash run/test_Pascal3D.sh

You should obtain the results in Table 2 in the paper (*indicates category-agnostic):

Method	Accuracy	Median Error
Keypoints and Viewpoints	80.75	13.6
Render for CNN	82.00	11.7
Mousavian	81.03	11.1
Grabner	83.92	10.9
Grabner*	81.33	11.5
StarMap*	81.67	12.8
Ours(MV)*	82.66	10.0

Pix3D

bash run/test_Pix3D.sh

You should obtain the results in Table 3 in the paper (Accuracy / MedErr):

Method	Bed	Chair	Desk
Georgakis	50.8 / 28.6	31.2 / 57.3	34.9 / 51.6
Ours(MV)	59.8 / 20.0	52.4 / 26.6	56.6 / 26.6

Demo

In order to test on other 3D model, first you need to generate multiviews from .obj file by running python ./data/render_utils.py with the correct path and you should save the testing images picturing this model in a folder.

Then you can run bash ./demo/inference.sh to get predictions and images rendered under the predicted pose with the right model_path, image_path, render_path, obj_path.

Some example of applying our model trained on objects of ObjectNet3D with keypoint annotations to armadillo images can be seen below: