/PFNet-FGVC

PFNet: A Novel Part Fusion Network for Fine-grained Visual Categorization

Primary LanguagePython

Codes for PFNet: A Novel Part Fusion Network for Fine-grained Visual Categorization

This repository holds the PyTorch(V0.3.0) code for PFNet.

Introduction

The existing methods in fine-grained visual categorization focus on integrating multiple deep CNN models or complicated attention mechanism, resulting in increasing cumbersome networks. In addition, most methods rely on part annotations which requires expensive expert guidance. In this paper, without extra annotation, we propose a novel part fusion network (PFNet) to effectively fuse discriminative image parts for classification. More specifically, PFNet consists of a part feature extractor to extract part features and a two-level classification network to utilize part-level and image-level features simultaneously. Part-level features are trained with the weighted part loss, which embeds a weighting mechanism based on different parts' characteristics. Easy parts, hard parts and background parts are proposed and discriminatively used for classification. Moreover, part-level features are fused to form an image-level feature so as to introduce global supervision and generate final predictions. Experiments on three popular benchmark datasets show that our framework achieves competitive performance compared with the state-of-the-art.

alt text

Prepare Datasets

Prepare the corresponding datasets (CUB-200-2011, Stanford Cars or FGVC-Aircraft) before training PFNet. For quick start, you can download the dataset Stanford Cars, proposed rois files car_rois500.tar.gz and prepared train/test split file car_splits.tar.gz. Unzip these files and organize them in the current working directory as follows:

-car
--car_ims
---000001.jpg

--car_rois500
---car_ims
----000001.txt

--split
---Acura Integra Type R 2001_test.txt

For part proposal, we also provide codes for generating part proposals using Selective Search Window. Please refer to the guide provide in our part proposal directory.

Usage

1, Download this repo recursively:

git clone --recursive https://github.com/MichaelLiang12/PFNet-FGVC.git

2, Build RoiPooling module

Please follow the instuctions in pytorch-faster-rcnn. We use the RoiPooling module implemented by them. Note that if you also use Ubuntu14.04+Cuda8.0+TitanX, you might not need to compile again.

3, Run PFNet_train_test.py

You can modify fundamental parameters in the main() function. The training process might be like follows. By setting args.evaluate = True, you can download our model and test it directly.

alt text

Citation

For Selective Search Window and RoiPooling module.

@article{uijlings2013selective,
  title={Selective search for object recognition},
  author={Uijlings, Jasper RR and Van De Sande, Koen EA and Gevers, Theo and Smeulders, Arnold WM},
  journal={International Journal of Computer Vision},
  volume={104},
  number={2},
  pages={154--171},
  year={2013},
  publisher={Springer}
}

@article{chen17implementation,
    Author = {Xinlei Chen and Abhinav Gupta},
    Title = {An Implementation of Faster RCNN with Study for Region Sampling},
    Journal = {arXiv preprint arXiv:1702.02138},
    Year = {2017}
}

Citation for our PFNet

@Article{Liang2018,
author="Liang, Jingyun
and Guo, Jinlin
and Guo, Yanming
and Lao, Songyang",
title="PFNet: a novel part fusion network for fine-grained visual categorization",
journal="Multimedia Tools and Applications",
year="2018",
month="Dec",
day="15",
issn="1573-7721",
doi="10.1007/s11042-018-7047-5",
url="https://doi.org/10.1007/s11042-018-7047-5"
}

View Paper