myfamily111 / hardnet_multidataset Goto Github PK

Training HardNet8 on multiple datasets

Provided code allows to train HardNet8 (described in the thesis) on multiple datasets of format of Liberty or AMOS. Use the script ftrain.py for training.

Pretrained models are available in ".pt" and ".jitpt" formats. The latter can be loaded via torch.jit.load(name) without any additional code.

If you want to train the model, you must first download and unzip the datasets. Please make sure to use the Liberty dataset from this file because it contains also source image IDs.

Example:

Run this to train the universal HardNet8-Univ descriptor on AMOS and Liberty datasets.

python -utt ftrain.py --arch=h8 --ds=v4+lib --loss=tripletMargin++

This command was used to train HardNet8-PTv2 submitted to CVPR IMW 2020.

python -utt ftrain.py --arch=h8E512 --ds=lib+colo+notre --bs=9000 --mpos=0.5 --fewcams

In both cases, to get the final model, run the code in Notebooks/add_pca.ipynb (you have to change the name of model) to create new checkpoint with added PCA compression.

Please cite us if you use this code:

Leveraging Outdoor Webcams for Local Descriptor Learning

@article{HardNetAMOS2019,
    author 	= {Milan Pultar, Dmytro Mishkin, Jiri Matas},
    title  	= "{Leveraging Outdoor Webcams for Local Descriptor Learning}",
    year   	= 2019,
    month    	= feb,
    booktitle 	= {Proceedings of CVWW 2019}
}

Working hard to know your neighbor's margins: Local descriptor learning loss

@article{HardNet2017,
    author 	= {Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic, Jiri Matas},
    title 	= "{Working hard to know your neighbor's margins: Local descriptor learning loss}",
    year 	= 2017}
(c) 2017 by Anastasiia Mishchuk, Dmytro Mishkin