lejeunel / cobnet

convolution oriented boundaries (pytorch)

TODO

• Integrate Higra to generate multiscale hierarchy
• Check training of orientation maps
• Check results with original caffe implementation

Description

PyTorch implementation of Convolutional Oriented Boundaries

Differences w.r.t original implementation:

• ResNet50 from pyTorch model zoo, which differs from author's Caffe model (has batch normalization layers)
• Batch size of 16
• Base learning-rate is 1e-4 and is increased for "deeper" layers
• Weight initialization is gaussian/normal instead of constant

Dependencies

Most of these are easily installed with your favorite package manager

• PyTorch >= 1.0
• Numpy
• Scipy
• imgaug
• opencv-python
• tqdm
• imgaug
• sklearn
• tensorboardX
• higra

Dataset

Download and uncompress the following datasets/annotations:

• Pascal VOC 2012 Unzip this in /pascal-voc
• Pascal-Context Unzip this in /trainval

Training

On first run, the whole dataset (>10k images) will be processed to extract boundaries, this can take more than an hour!

python train.py --root-imgs <root>/pascal-voc/VOC2012 --root-segs <root>/trainval --run-path <root>/runs/cob --cuda