Locating Objects Without Bounding Boxes

PyTorch code for "Locating Objects Without Bounding Boxes" , CVPR 2019 - Oral, Best Paper Finalist (Top 1 %) [Paper] [Youtube]

Citing this work

@article{ribera2019,
  title={Locating Objects Without Bounding Boxes},
  author={Javier Ribera and David G\"{u}era and Yuhao Chen and Edward J. Delp},
  journal={Proceedings of the Computer Vision and Pattern Recognition (CVPR)},
  month={June},
  year={2019},
  note={{Long Beach, CA}}
}

Datasets

The datasets used in the paper can be downloaded from:

• Mall dataset
• Pupil dataset
• Plant dataset

Installation

Use conda to recreate the environment provided with the code:

conda env create -f environment.yml

Activate the environment:

conda activate object-locator

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 TRAINING_DIRECTORY \
       --batch-size 32 \
       --visdom-env mytrainingsession \
       --visdom-server localhost \
       --lr 1e-3 \
       --val-dir TRAINING_DIRECTORY \
       --optim Adam \
       --save saved_model.ckpt

Dataset format

The options --dataset and --train-dir should point to a directory. This directory must contain your dataset, meaning:

1. One file per image to analyze (png, jpg, jpeg, tiff or tif).
2. One ground truth file called gt.csv with the following format:

filename,count,locations
img1.png,3,"[(28, 52), (58, 53), (135, 50)]"
img2.png,2,"[(92, 47), (33, 82)]"

Each row of the CSV must describe the ground truth of an image: the count (number) and location of all objects in that image. The locations are in (y, x) format, being the origin the most top left pixel, y being the pixel row number, and x being the pixel column number.

Optionally, if you are working on precision agriculture or plant phenotyping you can use an XML file gt.xml instead of a CSV. The required XML specifications can be found in https://communityhub.purdue.edu/groups/phenosorg/wiki/APIspecs (accessible only to Purdue users) and in this thesis, but this is only useful in agronomy/phenotyping applications. The XML file is parsed by the file data_plant_stuff.py.

Pre-trained models

Models are trained separately for each of the four datasets, as described in the paper:

1. Mall dataset
2. Pupil dataset
3. Plant dataset
4. ShanghaiTechB dataset

The COPYRIGHT of the pre-trained models is the same as in this repository.

As described in the paper, the pre-trained model for the pupil dataset excludes the five central layers. Thus if you want to use this model you will have to use the option --ultrasmallnet.

Uninstall

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

Code Versioning

The code used in the paper corresponds to the tag used-for-cvpr2019-submission. If you want to reproduce the results, checkout that tag with git checkout used-for-cvpr2019-submission. The master branch is the latest version available, with convenient bug fixes and better documentation. If you want to develop or retrain your models, we recommend the master branch. Versions numbers follow semantic versioning and the changelog is in CHANGELOG.md.