This is the official implementation of our work entitled as Universal Fine-Grained Visual Categorization by Concept Guided Learning, which has been accepted by IEEE Transactions on Image Processing (TIP'2025).
-
For
GoogleDriveuser, please click here to download. -
For
BaiduDriveuser, please click here to download.
Million-AID dataset has about one million aerial scene samples from high-resolution satellite images from a variety of imaging sensors (e.g., worldview-2, Gaofen-2, and etc.) in total, but the weakness includes:
-
Most of the samples in million-AID are annotated automatically or semi-automatically, i.e., without human-level supervision.
-
Only about 10,000 samples have the publicly-available ground truth, which poses a bottleneck on the amount of training data.
In this work, the proposed FGLCD makes the following advancement compared with the previous Million-AID:
-
manually select and correct the annotation of the samples
-
enlarge the size of benchmark: a total of 59994 samples (29998 for training, 29996 for testing)
-
The first dataset for the task of fine-grained land-cover scene classification. Different from conventional remote sensing scene classification datasets, such as
UCM,AIDandNWPU, the fine-grained categorization strictly follows the land-use classification standards GB/T 21010-2017. -
A total of 51 geo-spatial fine-grained categories from 8 coarse-grained categories.
-
A total of 59994 samples (29998 for training, 29996 for testing).
To set up the environment, please install the following packages:
matplotlib==3.3.1
numpy==1.20.2
opencv-python==4.5.2.54
pillow==8.2.0
pip==21.1.3
seaborn==0.11.0
timm==0.5.4
torch==1.9.0
torchvision==0.10.0
wandb==0.12.4
The training and inference command is:
python main.py --c configs/MTARSI_SwinT.yaml
Please remember to change the file folder to your own in the .yaml file.
The development of CGL largely relies on the source code from FGVC-PIM, with the code link [https://github.com/chou141253/FGVC-PIM]. We sincerely appreciate the authors of A Novel Plug-in Module for Fine-grained Visual Classification to advance fine-grained visual categorization.
If you find this work useful for your research, please cite our work as follows:
@article{bi2025universal,
title={Universal Fine-grained Visual Categorization by Concept Guided Learning},
author={Bi, Qi and Zhou, Beichen and Ji, Wei and Xia, Gui-Song},
journal={IEEE Transactions on Image Processing},
volume={34},
pages={394--409},
year={2025}
}