This is an simple implementation of Multi-Objective Matrix Normalization for Fine-grained Visual Recognition by Pytorch. Paper is accepted by TIP. The code will be re-organized to make it more clear.
pytorch 1.0
The training scripts for CUB, Cars, Air, and Dogs are given in https://drive.google.com/drive/folders/1mgKoXwDg3oUGiJluCSWlZJkvrhsbq2tw?usp=sharing. Other extensions can be easily modified.
A detailed illustration is as follows:
step 1:
adding your data path around the 130 line in main.py
step 2:
creating a running bash scrip, as the given example in the google drive. specifically, the running command should be given by:
python main.py -a momn -d cub -s ./cub/checkpoints --backbone densenet201 -b 230 --lr 0.1 --resize_size 560 --crop_size 512 --epochs 90 --is_fix --pretrained
An example script for testing is also given in the google drive.
Several pretrained models are given in https://drive.google.com/drive/folders/1mgKoXwDg3oUGiJluCSWlZJkvrhsbq2tw?usp=sharing, of which the performance is listed in the paper.