Robust Learning from Noisy Web Data for Fine-Grained Recognition

Introduction

This is the PyTorch implementation for our paper Robust Learning from Noisy Web Data for Fine-Grained Recognition

Network Architecture

The architecture of our proposed approach is as follows

Environment

Create a virtual environment with python 3.7,

$  conda create -n msdejor_env python=3.7

$  conda activate msdejor_env

Install all dependencies

$  pip install -r requirements.txt

Data Preparation

Download these web fine-grained datasets, namely Web-CUB, Web-Car and Web-Aircraft. Then uncompress them into ./data directory.

---data
   ├── web-bird
   │   ├── train
   │   └── val
   ├── web-car
   │   ├── train
   │   └── val
   └── web-aircraft
       ├── train
       └── tval

Training

If you want to use multi-scale module, modify the corresponding parameters in main_msdejor.py or directly run main_msdejor.py to get the final result. We provide the default parameter settings as following：

python main_msdejor.py --bs 30 --net 50 --data bird --lamb 0.1 --gama 2

If you only prefer the DeJoR module, run main_dejor.py.

python main_dejor.py --bs 50 --net 18 --data bird --lamb 0.1 --gama 2

In our experiments, we adopt the same hyperparameters across three benchmark datasets, setting to 0.1 and to 2.

The final experimental results are shown in the following table：

Codeczh/MS-DeJOR