https://github.com/zhaoxin94/fsac

https://github.com/zhaoxin94/fsac

Last synced: 4 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/zhaoxin94/fsac
• Owner: zhaoxin94
• License: mit
• Created: 2022-07-30T03:43:58.000Z (almost 4 years ago)
• Default Branch: main
• Last Pushed: 2021-12-16T05:44:44.000Z (over 4 years ago)
• Last Synced: 2025-02-28T09:11:37.382Z (over 1 year ago)
• Homepage:
• Size: 7.67 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Frequency Spectrum Augmentation Consistency for Domain Adaptive Object Detection

## Main requirements

torch >= 1.0

torchvision >= 0.2.0

Python 3

## Environmental settings

This repository is developed using python 3.6.12 on Ubuntu 16.04.5 LTS. The CUDA and pytorch version is 11.2 and 1.7.1. We use one NVIDIA 3090 GPU card for training and testing.

## Dataset

PASCAL VOC, Watercolor, Cityscapes, Foggycityscapes -> Please follow the instructions in [[Link](https://github.com/VisionLearningGroup/DA_Detection)] to prepare the datasets.

Daytime-Sunny, Dusk-Rainy, and Night-Rainy -> Dataset preparation instruction link [[Link](https://github.com/AmingWu/VDD-DAOD)].

## Code

Faster R-CNN -> Thanks for jwyang [[Link](https://github.com/jwyang/faster-rcnn.pytorch/tree/pytorch-1.0)]; Fourier Domain Adaptation -> Thanks for Yanchao Yang [[Link](https://github.com/YanchaoYang/FDA)].

Our Augmentation (Mix+Replace+Extend+Disorder).

## Train

To train a faster R-CNN model with vgg16 on pascal_voc:

```

CUDA_VISIBLE_DEVICES=$GPU_ID python trainval_net.py --dataset pascal_voc --net vgg16 --bs 1 --cuda

```

And you need to add augmentated data in the loadpath by creating a new dataset_name variable.

## Test

To test:

```

python test_net.py --dataset pascal_voc --net vgg16 --modelpath your modelpath --cuda

```

## Augmentation

Daytime-Sunny -> Dusk-Rainy

![shapenet_illuminants](image/aug1.png)

Daytime-Sunny -> Night-Rainy

![shapenet_illuminants](image/aug2.png)

## Result

![shapenet_illuminants](image/result.png)

Results on adaptation from Cityscapes to FoggyCityscapes. ‘prsn’, ‘mcycl’, and ‘bcycl’ separately denote ‘person’, ‘motorcycle’, and ‘bicycle’ category.

![shapenet_illuminants](image/table2.png)

Results on adaptation from Daytime-sunny to Duskrainy. Here, we directly run the released codes of the compared methods to obtain the results.

![shapenet_illuminants](image/table3.png)

Results on Daytime-sunny → Night-rainy.

![shapenet_illuminants](image/table4.png)

Results on the compound target domain.

![shapenet_illuminants](image/table5.png)