开源软件名称（OpenSource Name）：

开源软件地址(OpenSource Url)：

开源编程语言(OpenSource Language)：

开源软件介绍(OpenSource Introduction)：

A loss function (Weighted Hausdorff Distance)
for object localization

This repository contains the PyTorch implementation of the Weighted Hausdorff Loss described in this paper: Weighted Hausdorff Distance: A Loss Function For Object Localization

Abstract

Recent advances in Convolutional Neural Networks (CNN) have achieved remarkable results in localizing objects in images. In these networks, the training procedure usually requires providing bounding boxes or the maximum number of expected objects. In this paper, we address the task of estimating object locations without annotated bounding boxes, which are typically hand-drawn and time consuming to label. We propose a loss function that can be used in any Fully Convolutional Network (FCN) to estimate object locations. This loss function is a modification of the Average Hausdorff Distance between two unordered sets of points. The proposed method does not require one to "guess" the maximum number of objects in the image, and has no notion of bounding boxes, region proposals, or sliding windows. We evaluate our method with three datasets designed to locate people's heads, pupil centers and plant centers. We report an average precision and recall of 94% for the three datasets, and an average location error of 6 pixels in 256x256 images.

Citation

J. Ribera, D. Güera, Y. Chen, E. Delp, "Weighted Hausdorff Distance: A Loss Function For Object Localization", arXiv preprint arXiv:1806.07564, June 2018

@article{whd-loss,
  title={Weighted Hausdorff Distance: A Loss Function For Object Localization},
  author={J. Ribera and D. G{\"u}era and Y. Chen and E. Delp},
  journal={arXiv:1806.07564},
  month={June},
  year={2018}
}

Examples

Datasets

The datasets used in the paper can be downloaded from these links:

• Mall dataset
• Pupil dataset
• Plant dataset

Code

The code used for the Arxiv submission corresponds to the tag used-for-arxiv-submission. If you wish to reproduce the results, checkout that tag with git checkout used-for-arxiv-submission. The master branch is the latest version available.

Installation

Use conda to recreate the environment provided with the code:

conda env create -f environment.yml

and install the tool:

pip install .

Usage

Activate the environment:

conda activate object-locator

Run this to get help (usage instructions):

python -m object-locator.locate -h
python -m object-locator.train -h

Example:

python -m object-locator.locate \
       --dataset DIRECTORY \
       --out DIRECTORY \
       --model CHECKPOINTS \
       --evaluate \
       --no-gpu \
       --radius 5

python -m object-locator.train \
       --train-dir ~/data/20160613_F54_training_256x256 \
       --batch-size 32 \
       --env-name sorghum \
       --lr 1e-3 \
       --val-dir ~/data/plant_counts_random_patches/20160613_F54_validation_256x256 \
       --optim Adam \
       --save unet_model.ckpt

Uninstall

conda deactivate object-locator
conda env remove --name object-locator