Mesh R-CNN with Sheared Decoder

Advanced Deep Learning for Computer Vision - Visual Computing

We propose a novel non-learnable Shearing Layer, that can lift 2D feature maps to 3D feature maps in a computationally efficient manner, without the need for backprojection. We have implemented the same on top of the Mesh-RCNN code base.

Installation Requirements

• [Detectron2][d2]
• [PyTorch3D][py3d]

The implementation of Mesh R-CNN is based on [Detectron2][d2] and [PyTorch3D][py3d]. You will first need to install those in order to be able to run Mesh R-CNN.

To install

git clone https://github.com/facebookresearch/meshrcnn.git
cd meshrcnn && pip install -e .

Demo

Run Mesh R-CNN with Sheared Decoder on an input image

python demo/demo.py \
--config-file configs/pix3d/meshrcnn_R50_FPN.yaml \
--input /path/to/image \
--output output_demo \
--onlyhighest MODEL.WEIGHTS meshrcnn://meshrcnn_R50.pth

See demo.py for more details.

Running Experiments

Pix3D

See INSTRUCTIONS_PIX3D.md for more instructions.