Introduction

FGVCLib is an open-source and well documented library for Fine-grained Visual Classification. It is based on Pytorch with performance and friendly API. Our code is pythonic, and the design is consistent with torchvision. You can easily develop new algorithms, or readily apply existing algorithms. The branch works with torch 1.12.1, torchvision 0.13.1.

Major features

• Modular Design

  We decompose the detection framework into different components and one can easily construct a customized object detection framework by combining different modules.

• State of the art We implement state-of-the-art methods by the FGVCLib, PMG, MCL, API-Net, CAL, TransFG, PIM.

Installation

Please refer to Installation for installation instructions.

Getting started

Please see get_started.md for the basic usage of FGVCLib. We provide the tutorials for:

• with existing data existing model
• with existing data new model
• learn about apis
• learn about configs
• learn about criterions
• learn about datasets
• learn about metrics
• learn about model
• learn about transforms
• learn about the tools

Overview of Benchmark and Model Zoo

Architectures

Fine-grained Visual Classification	Other
Baseline_ResNet50 Mutual-Channel-Loss PMG-ResNet50 PMG_V2_ResNet50 API-Net CAL TransFG PIM	visualization

Components

Backbones	Encoders	Heads	Necks	Sotas
Resnet VGG	Global Max Pooling Global Avg Pooling Max Pooling 2d	Classifier_1_FC Classifier_2_FC	Multi-scale Convolution neck	Baseline_ResNet50 Mutual-Channel-Loss PMG-ResNet50 PMG_V2_ResNet50 API-Net CAL TransFG PIM

The Result of the SOTA

We used fgvclib to replicate the state-of-the-art model, and the following table shows the results of our experiment.

SOTA	Result of the paper	Result of the official code	Result of the FGVCLib
API-Net	88.1	87.2	86.8
CAL	90.6	89.6	89.5
TransFG	91.7	91.1	89.3
PIM	92.8	91.9	91.4

Contact

Thanks for your attention! If you have any suggestion or question, you can leave a message here or contact us directly:

• [email protected]
• [email protected]

Others

In addition to fgvclib, we have developed a practical mini program on fine-grained visual classification, which can be accessed by scanning the two-dimensional code below.