philkr / lpo

Implementation of the CVPR 2015 paper: Learning to propose objects

Learning to propose objects

This implements the learning and inference/proposal algorithm described in "Learning to Propose Objects, Krähenbühl and Koltun, CVPR 2015".

Dependencies:

• c++11 compiler (gcc >= 4.7)
• cmake
• boost-python
• python (2.7 or 3.1+ should both work)
• numpy
• libmatio (optional)
• libpng, libjpeg
• Eigen 3 (3.2.0 or newer)
• OpenMP (optional but recommended)

Compilation:

Go to the top level directory

mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DDATA_DIR=/path/to/datasets -DUSE_PYTHON=ON
make -j9

Here "-DUSE_PYTHON" specifies that the python wrapper should be built (highly recommended). You can use python 2.7 by specifying "-DUSE_PYTHON=2", any other argument will try to build a python 3 wrapper.

The flag "-DDATA_DIR=/path/to/datasets" is optional and can point to a directory containing the VOC2012, VOC2007 or COCO datset. Specify this path if you want to train or evaluate LPO on those dataset.

"/path/to/datasets" can be any directory containing subdirectories:

• 'VOC2012/ImageSets'
• 'VOC2012/SegmentationClass',
• 'VOC2012/Annotations'
• 'COCO/train2014'
• 'COCO/val2014'
• ...

and files:

• 'COCO/instances_train2014.json'
• 'COCO/instances_val2014.json'.

The coco files can be downloaded from http://mscoco.org/, the PASCAL VOC dataset http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2012/index.html .

The code should compile and run fine on both Linux and Mac OS, let me know if you have any difficulty or find a bug. For Windows you're on your own.

Experiments

The code to reproduce most results in the paper is included here. All experiments should be run from the src directory.

To generate the main comparison in table 3 run:

bash eval_all.sh

To analyze a model like table 2 run:

python analyze_model.py path/to/model

To do the bounding box evaluation call:

python eval_box.py path/to/output_file path/to/model1 path/to/model2 path/to/model3 path/to/model4

This will create a binary file measuring number of proposals vs best overlap per object. You can then use the results/box.py script to generate the bounding box evaluation and produce the plots. For your convenience we included the precomputed results of many prior methods on VOC 2012 in results/box/*.dat.

Citation

If you're using this code in a scientific publication please cite:

@inproceedings{kk-lpo-15,
  author    = {Philipp Kr{\"{a}}henb{\"{u}}hl and
               Vladlen Koltun},
  title     = {Learning to Propose Objects},
  booktitle = {CVPR},
  year      = {2015},
}

License

All my code is published under a BSD license, so feel free to reuse and/or share it. There are some dependencies which are under different licenses and/or patented. All those dependencies are located in the external directory.