Shelf-supervised Mesh Prediction in the wild

in CVPR 2021, Yufei Ye, Shubham Tulsiani, Abhinav Gupta

Project Page, Video, Arxiv

We aim to infer 3D shape and pose from a single image and are able to train the system with only image collecitons and segmentation -- no template, camera pose, or multi-view association. The method consists of 2 steps:

1. Category-level Reconstruction. We first infer a volumetric representation in a canonical frame, along with the camera pose for the input image.
2. Instance-level Specialization. The coarse volumetric prediction is converted to a mesh-based representation, which is further optimized in the predicted camera frame given the input image.

This code repo is a re-implementation of the paper. The code is developed based on Pytorch 1.3 (Pytorch >=1.5 adds backprop version check which will trigger a runtime error), Pytorch3d 0.2.0, and integrated LPIPS. To voxelize meshes for evaluation, we use util code in Occupancy Net but did not include it in this reimplementation.

Demo: Estimate mesh with our pretrained model

Download pretrained models to weights/

python demo.py  --checkpoint=weights/wildchair.pth

Similar results should be saved at outputs/

input	output shape	output shape w/ texture

or for other curated categories:

python demo.py  --checkpoint=weights/cub.pth --demo_image examples/cub_0.png
python demo.py  --checkpoint=weights/wildchair.pth --demo_image examples/wildchair_0.png
python demo.py  --checkpoint=weights/quad.pth --demo_image examples/llama.png

for openimages 50 categories, the following script will reconstruct images under data/demo_images/:

python demo_all_cls.py 

Training

To train your own model, set up dataset following dataset.md before running

python train_test.py     --dataset allChair --cfg_file config/pmBigChair.json 

For more training details, please refer to train.md

Citation

If you find this work useful, please consider citing:

@inProceedings{ye2021shelf,
  title={Shelf-Supervised Mesh Prediction in the Wild},
  author={Ye, Yufei and Tulsiani, Shubham and  Gupta, Abhinav},
  year={2021},
  booktitle={Computer Vision and Pattern Recognition (CVPR)}
}