Pixel Attentional Gating for Parsimonious Per-Pixel Labeling
Switch branches/tags
Nothing to show
Clone or download
Latest commit 1bad30b Nov 10, 2018
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
demo1_Panoramic_Surface_Normal readme for demo1 Jun 29, 2018
libs demo1 Jun 29, 2018
matconvnet-1.0-beta23_modifiedDagnn demo1 Jun 29, 2018
README.md Update README.md Nov 9, 2018

README.md

Pixel-wise Attentional Gating for Scene Parsing

For paper and slides, please refer to our project page

Our entry to Robust Vision Challenge can be found here depth estimation and semantic segmentation.

alt text

To achieve parsimonious inference in per-pixel labeling tasks with a limited computational budget, we propose a Pixel-wise Attentional Gating unit (PAG) that learns to selectively process a subset of spatial locations at each layer of a deep convolutional network. PAG is a generic, architecture-independent, problem-agnostic mechanism that can be readily ``plugged in'' to an existing model with fine-tuning. We utilize PAG in two ways: 1) learning spatially varying pooling fields that improve model performance without the extra computation cost associated with multi-scale pooling, and 2) learning a dynamic computation policy for each pixel to decrease total computation while maintaining accuracy.

We extensively evaluate PAG on a variety of per-pixel labeling tasks, including semantic segmentation, boundary detection, monocular depth and surface normal estimation. We demonstrate that PAG allows competitive or state-of-the-art performance on these tasks. Our experiments show that PAG learns dynamic spatial allocation of computation over the input image which provides better performance trade-offs compared to related approaches (e.g., truncating deep models or dynamically skipping whole layers). Generally, we observe PAG can reduce computation by 10% without noticeable loss in accuracy and performance degrades gracefully when imposing stronger computational constraints.

Keywords Spatial Attention, Dynamic Computation, Per-Pixel Labeling, Semantic Segmentation, Monocular Depth, Surface Normal, Boundary Detection.

Several demos are included as below. As for details on the training, demo and code, please go into each demo folder.

  1. demo1: Panoramic Surface Normal Estimation [Ready]

  2. demo2: Boundary Detection [[!!!TOOD!!!]]

  3. demo3: Semantic Segmentation [[!!!TOOD!!!]]

  4. demo4: Monocular Depth Estimation [[!!!TOOD!!!]]

Please download those models from the google drive.

MatConvNet is used in our project, and some functions are changed/added. Please compile accordingly by adjusting the path --

LD_LIBRARY_PATH=/usr/local/cuda/lib64:local matlab 

path_to_matconvnet = './matconvnet-1.0-beta23_modifiedDagnn/';
run(fullfile(path_to_matconvnet, 'matlab', 'vl_setupnn'));
addpath(fullfile(path_to_matconvnet, 'matlab'));
vl_compilenn('enableGpu', true, ...
               'cudaRoot', '/usr/local/cuda', ...
               'cudaMethod', 'nvcc', ...
               'enableCudnn', true, ...
               'cudnnRoot', '/usr/local/cuda/cudnn/lib64') ;

See also Recurrent Scene Parsing with Perspective Understanding in-the Loop which adapts depth map for pooling field selection.

If you find our model/method/dataset useful, please cite our work (draft at arxiv):

@inproceedings{kong2019pag,
  title={Pixel-wise Attentional Gating for Scene parsing},
  author={Kong, Shu and Fowlkes, Charless},
  booktitle={IEEE Winter Conf. on Applications of Computer Vision (WACV)},
  year={2019}
}

last update: 11/06/2018

Shu Kong

aimerykong At g-m-a-i-l dot com

pag4ppl@gmail.com