ML-GCN.pytorch

PyTorch implementation of Multi-Label Image Recognition with Graph Convolutional Networks, CVPR 2019.

Update

1. In our original conference paper, we report the baseline classification results using GAP for comparison, because GAP is the default choice for feature aggregation in ResNet series. In our experiments, we found that replacing GAP with GMP leads to improved performance, and thus adopt GMP with our GCN method -- we regard GMP as one part of our method. For clarification, we re-run the baselines and here report the corresponding results in the following table.

Method	COCO	NUS-WIDE	VOC2007
Res-101 GAP	77.3	56.9	91.7
Res-101 GMP	81.9	59.7	93.0
Ours	83.0	62.5	94.0

2. We correct the typos in Eq. (8) as follows.

Requirements

Please, install the following packages

• numpy
• torch-0.3.1
• torchnet
• torchvision-0.2.0
• tqdm

Download pretrain models

checkpoint/coco (GoogleDrive)

checkpoint/voc (GoogleDrive)

or

Baidu

Options

• lr: learning rate
• lrp: factor for learning rate of pretrained layers. The learning rate of the pretrained layers is lr * lrp
• batch-size: number of images per batch
• image-size: size of the image
• epochs: number of training epochs
• evaluate: evaluate model on validation set
• resume: path to checkpoint

Demo VOC 2007

python3 demo_voc2007_gcn.py data/voc --image-size 448 --batch-size 32 -e --resume checkpoint/voc/voc_checkpoint.pth.tar

Demo COCO 2014

python3 demo_coco_gcn.py data/coco --image-size 448 --batch-size 32 -e --resume checkpoint/coco/coco_checkpoint.pth.tar

Citing this repository

If you find this code useful in your research, please consider citing us:

@inproceedings{ML-GCN_CVPR_2019,
author = {Zhao-Min Chen and Xiu-Shen Wei and Peng Wang and Yanwen Guo},
title = {{Multi-Label Image Recognition with Graph Convolutional Networks}},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2019}
}

Reference

This project is based on https://github.com/durandtibo/wildcat.pytorch

Tips

If you have any questions about our work, please do not hesitate to contact us by emails.