Skip to content
Go to file
This branch is 39 commits ahead, 4 commits behind rbgirshick:master.

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

PVANet: Lightweight Deep Neural Networks for Real-time Object Detection

by Sanghoon Hong, Byungseok Roh, Kye-hyeon Kim, Yeongjae Cheon, Minje Park (Intel Imaging and Camera Technology) Presented in EMDNN2016, a NIPS2016 workshop (arXiv link)


This repository is a fork from py-faster-rcnn and demonstrates the performance of PVANet.

You can refer to py-faster-rcnn and faster-rcnn for more information.


Please note that this repository doesn't contain our in-house codes used in the published article.

  • This version of py-faster-rcnn is slower than our in-house runtime code (e.g. image pre-processing code written in Python)
  • PVANet was trained by our in-house deep learning library, not by this implementation.
  • There might be a tiny difference in VOC2012 test results, because some hidden parameters in py-faster-rcnn may be set differently with ours.

Citing PVANet

If you want to cite this work in your publication:

  title={{PVANet}: Lightweight Deep Neural Networks for Real-time Object Detection},
  author={Hong, Sanghoon and Roh, Byungseok and Kim, Kye-Hyeon and Cheon, Yeongjae and Park, Minje},
  journal={arXiv preprint arXiv:1611.08588},


  1. Install Dependency

       sudo pip install Cython
  2. Clone the Faster R-CNN repository

    # Make sure to clone with --recursive
    git clone --recursive
  3. We'll call the directory that you cloned Faster R-CNN into FRCN_ROOT. Build the Cython modules

    cd $FRCN_ROOT/lib
  4. Build Caffe and pycaffe

    cd $FRCN_ROOT/caffe-fast-rcnn
    # Now follow the Caffe installation instructions here:
    # For your Makefile.config:
    #   Uncomment `WITH_PYTHON_LAYER := 1`
    cp Makefile.config.example Makefile.config
    make -j8 && make pycaffe
  5. Download PVANet detection model for VOC2007

    cd $FRCN_ROOT
  6. Download PVANet detection model for VOC2012 (published model)

    cd $FRCN_ROOT
  7. (Optional) Download all available models (including pre-trained and compressed models)

    cd $FRCN_ROOT
  8. (Optional) Download ILSVRC2012 (ImageNet) classification model

    cd $FRCN_ROOT
  9. (Optional) If the scripts don't work, please download the models from ...

    Model Google Drive
    PVANet for VOC2007 link
    PVANet for VOC2012 link
    PVANet for VOC2012 (compressed) link
    PVANet for ILSVRC2012 (ImageNet) link
    PVANet pre-trained link

How to run the demo

  1. Download PVANet detection models

  2. Run the demo script

    cd $FRCN_ROOT

How to test on the VOC dataset

  1. Download PASCAL VOC 2007 and 2012 -- Follow the instructions in py-faster-rcnn

  2. PVANet on PASCAL VOC 2007

    cd $FRCN_ROOT
    ./tools/ --net models/pvanet/pva9.1/PVA9.1_ImgNet_COCO_VOC0712.caffemodel --def models/pvanet/pva9.1/ --cfg models/pvanet/cfgs/submit_1019.yml --gpu 0
  3. PVANet (compressed)

    cd $FRCN_ROOT
    ./tools/ --net models/pvanet/pva9.1/PVA9.1_ImgNet_COCO_VOC0712plus_compressed.caffemodel --def models/pvanet/pva9.1/ --cfg models/pvanet/cfgs/submit_1019.yml --gpu 0

Expected results

Mean Average Precision on VOC detection tasks

Model VOC2007 mAP (%) VOC2012 mAP (%)
PVANet+ (VOC2007) 84.9 N/A
PVANet+ (VOC2012) 89.8 84.2
PVANet+ (VOC2012 + compressed) 87.8 83.7
  • The training set for the VOC2012 model includes the VOC2007 test set. Therefore the accuracies on VOC2007 of the model are not meaningful; They're shown here just for reference

Validation error on ILSVRC2012

Input size Top-1 error (%) Top-5 error (%)
192x192 30.00 N/A
224x224 27.66 8.84
  • We re-trained a 224x224 model from the '192x192' model as a base model.
You can’t perform that action at this time.