Landmark Feature Convolutional Networks

This repo implements landmarkconv that aims to learn convolutional features outside the box.

1. Standard conv has a quite limited receptive field.
2. Dilated conv enlarges the receptive field without introducing extra paramter, but leads to many holes.
3. Deformvable conv make the receptive field more flexible by adding learnable offsets. However, the receptive filed is still box-like since the learned offsets are usually small.

1. While box conv updates representations with neighboring points, we update representation with neighboring regions, which are extracted with a permutation invariant function.

2. We part the whole image into several intervals according to the given "kernel size" (number of intervals).

3. The landmark conv also enlarges the receptive filed without introducing extra paramter.

1. Spatial robustness
2. Long range information modeling

To understand, here is a very simple example (though not restrictly)

The landmark conv tends to generate equivalent representation if the they have the same spatial relationships. 🐷💗🐻

Visualization in natural image.

• Linux with Python 3.6
• CUDA support
• gcc 5.4

# git clone git://github.com/hbb1/landmarkconv
cd lanmarkconv/lib/layers
make

For convenience, LandmarkConvs are implemented as subclass of torch.nn.Conv2D, so just use it like standard convolution in Pytorch.

from .lib.layers.conv4 import PConv2d4
from .lib.layers.conv8 import PConv2d8

• Performance on visual grounding (image), see LBYLNet
• Visual grounding in other domain (video, 3D, RGBD).
• Object detection and semantic segmentation and so on.

The convolution is first proposed in this paper, if you find it helpful in your research, please cite our paper.

@InProceedings{huang2021look,
      title={Look Before You Leap: Learning Landmark Features for One-Stage Visual Grounding}, 
      author={Huang, Binbin and Lian, Dongze and Luo, Weixin and Gao, Shenghua},
      booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
      month = {June},
      year={2021},
}