TSCM Triplet Style-Content Metric

This repository includes introductions and implementation of Style-Content Metric Learning for Multidomain Remote Sensing Object Recognition in PyTorch.

Datasets

We conduct experiments using four remote sensing datasets: NWPU VHR-10 (Cheng, Zhou, and Han 2016), DOTA (Xia et al. 2018), HRRSD (Zhang et al. 2019) and DIOR (Li et al. 2020c).

Remote sensing objects are cut out from object detection ground truth.

Ten common object categories among four datasets are reserved for experiments, i.e., Airship, Ship, Storage Tank, Baseball Diamond, Tennis Court, Basketball Court, Ground Track Field, Harbor, Bridge and Vehicle .

Specifically, folder index and categories are as follows:

01 baseball_field
02 basketball_field
03 overpass
04 stadium
05 harbor
08 plane
10 ship
11 car
13 oil_tank
15 tennis_court

You can download post-processed datasets from these links( google drive ):
NWPU-RSOR,
DOTA-RSOR,
HRRSD-RSOR,
DIOR-RSOR.

File Structure

Trained model are saved in Ready Model folder, unzip ResNet-50-TSCM.zip.
Note that resnet50-pretrained.zip is a model provided by PyTorch pre-trained on ImageNet, we utilize it as initialized params of ResNet-50.

TSCM
├── data
│   ├── DIOR_RSOR
│   │   ├── test
│   │   ├── train
│   │   ├── eval
│   │   └── label_dict.txt
│   ├── DOTA_RSOR
│   ├── HRRSD_RSOR
│   └── NWPU_RSOR
├── label_dic_10.npy
├── Modules
├── Ready Model
│ ├── emvironment.yml
│ ├── label_dict.txt
│ ├── resnet50-pretrained.zip
│ └── ResNet-50-TSCM.zip
├── save
├── train_and_test.py
├── train.py
├── eval.py
└── version_check.py

Requirements

PyTorch >= 1.3.1
TorchVision >= 0.4.2
cv2 >= 3.4.2

Recommended
tqdm >= 4.61
apex >= 0.1

Train and Eval

Train

For detailed argparse params, run

python train.py -h

If you don't intend to customize params and paths, run

python train.py

Eval

For detailed argparse params, run

python eval.py -h

If you don't intend to customize params and paths, run

python eval.py