PyTorch implementation of Multi-Label Image Recognition with Graph Convolutional Networks, CVPR 2019.
- 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 |
- We correct the typos in Eq. (8) as follows.
Please, install the following packages
- numpy
- torch-0.3.1
- torchnet
- torchvision-0.2.0
- tqdm
checkpoint/coco (GoogleDrive)
checkpoint/voc (GoogleDrive)
or
lr: learning ratelrp: factor for learning rate of pretrained layers. The learning rate of the pretrained layers islr * lrpbatch-size: number of images per batchimage-size: size of the imageepochs: number of training epochsevaluate: evaluate model on validation setresume: path to checkpoint
python3 demo_voc2007_gcn.py data/voc --image-size 448 --batch-size 32 -e --resume checkpoint/voc/voc_checkpoint.pth.tarpython3 demo_coco_gcn.py data/coco --image-size 448 --batch-size 32 -e --resume checkpoint/coco/coco_checkpoint.pth.tarIf 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}
}
This project is based on https://github.com/durandtibo/wildcat.pytorch
If you have any questions about our work, please do not hesitate to contact us by emails.