SEF

This is the PyTorch implementation of our IEEE Signal Processing Letters paper "Learning Semantically Enhanced Feature for Fine-Grained Image Classification". We experimented on 4 fine-grained benchmark datasets --- CUB-200-2011, Stanford Cars, Stanford Dogs, and VGG-Aircraft. You should first download these datasets from their project homepages before runing SEF.

SEF can achieve comparable performance to the state-of-the-art methods with considerably less computation cost and training epochs. It has about 1.7% more parameters than its ResNet-50 backbone on the CUB-Birds dataset and can be easily integrated into other DCNNs.

Appoach

Requirements

PyTorch (>1.5),
CUDA 10.2
Python 3.6
Tensorboard (>2.2.2)

Learning and Evaluation

A "x-imdb.py" is provided for each dataset to generate Python pickle files, which are then used to prepare train/val/trainval/test data. Run "x-imdb.py" in the folder of your dataset to generate corresponding pickle file (imdb.pkl) should be the very first step.

main.py trains your own SEF model.
eval.py outputs classification accuracy by employing pretrained SEF models.

Due to the random generation of train/val/test data on some datasets, the classification accuracy may have a bit fluctuation but it should be in a reasonable range.

The pretrained SEF models can be download from HERE with code i5wk.

Results

Accuracy

SEF-18 and SEF-50 are results of models with different backbones (ResNet-18 and ResNet-50)

	SEF-18	SEF-50
CUB-200-2011	84.8%	87.3%
Stanford Cars	91.8%	94.0%
Stanford Dogs	83.1%	88.8%
VGG-Aircraft	89.3%	92.1%

Visualization

Citation

@inproceedings{sef@luowei,
author = {Wei Luo and Hengmin Zhang and Jun Li and Xiu-Shen Wei},
title = {Learning Semantically Enhanced Feature for Fine-Grained Image Classification},
booktitle = {arXiv preprint arXiv:2006.13457},
year = {2020},
}

youtang1993/SEF

SEF