This is the PyTorch implementation of our paper: Unsupervised underwater shipwreck detection in side-scan sonar images based on domain-adaptive techniques
| Methods on Cityscape to Foggy Cityscape | Backbone | density | mAP(%) |
|---|---|---|---|
| Source(Faster R-CNN) | Resnet-101 | 0.02 | 25.6 |
| Adaptive Teacher(CVPR2022) | Resnet-101 | 0.02 | 48.1 |
| DCBD(ours) | Resnet-101 | 0.02 | 52.6(+4.5) |
| Oracle | Resnet-101 | 0.02 | 43.2 |
| Methods on Optical to SSS Shipwreck | Backbone | mAP(%) |
|---|---|---|
| Source(Faster R-CNN) | Resnet-101 | 21.07 |
| CycleGAN(ICCV2017) | Resnet-101 | 50.37 |
| Adaptive Teacher(CVPR2022) | Resnet-101 | 73.47 |
| Adaptive Teacher + CycleGAN | Resnet-101 | 89.20 |
| DCBD(ours) | Resnet-101 | 92.16 |
| Oracle | Resnet-101 | 94.26 |
make sure you have set your dataset in /home/shu3090/wcw/adapteacher/data/datasets/builtin.py
train:
python train.py --config ./configs/faster_rcnn_R101_cross_city_res_change.yaml --num-gpus 2
resume:
python train.py --config ./configs/faster_rcnn_R101_cross_city_res_change.yaml --resume --num-gpus 2
eval-only:
python train.py --config ./configs/faster_rcnn_R101_cross_city_res_change.yaml --resume --eval-only
test:
python train.py --config ./configs/faster_rcnn_R101_cross_city_res_change.yaml --resume --eval-only
train DCBD:
#the first stage:
#train DID
python train_net.py --config ./configs/faster_rcnn_R101_cross_city_res_change.yaml --num-gpus 2
#train DRD
#1/Delete GRL module and IDCC module, retrain the detector
python train_net.py --config ./configs/faster_rcnn_R101_cross_city_res_change.yaml --num-gpus 2
#DCB:
#fusing precition from DID and DRD to train DAD, DS and DP means DID and DRD actually
python train_net_cb.py --config ./configs/faster_RCNN_city_cb.yaml --num-gpus 2 MODEL.WEIGHTS_DP="$your weight of DRD.pth$" MODEL.WEIGHTS_DS="$your weight of DID.pth$"
if you need a visualization demo for detection, We have implemented a simple demo using Gradio:
python visualization_demo.py --config ./configs/faster_rcnn_R101_cross_city_res_change.yaml
the dataset for optical to SSS task can be found at https://1drv.ms/f/c/02650644e5809154/ElSRgOVEBmUggAI3BwAAAAABGueJU3bLFGuZrcaFbeOgyQ?e=8bruXm