kamu1403 / meta-centernet-dior Goto Github PK

• This codebase is created to build benchmarks for few-shot learning object detection on DIOR dataset and provide a new model Meta-CenterNet.
• It is modified from mmfewshot.

• The loading module of DIOR dataset is provided in mmdetection module, and the configuration trained on DIOR dataset using CenternNet, Faster R-CNN, and Mask R-CNN is set up.
• The loading module of N-way K-shot DIOR dataset is provided in mmfewshot module, and four sample category segmentation is set according to the paper.
  In addition, the configuration that uses Meta-RCNN to train on DIOR dataset is set.
• A new model Meta-CenterNet is provided in mmfewshot module. Resnet and deconvolution network are used to extract feature map and supporting features. A new feature aggregation module CorrelationAggregator is used to convolve and combine the two features to obtain the aggregated feature map. And the detection box is extracted from the anchor-free detection head based on CenterNet. The configuration for training on DIOR dataset is set.

This project is released under the Apache 2.0 license.

Please refer to requirements.txt or follow the setup instruction of mmfewshot

MMFewShot depends on PyTorch and MMCV. Please refer to install.md for installation of MMFewShot

Download DIOR dataset and place it under dataset folder. It should be like

/mmfewshot/data/dior/
        Annnotations/
                Horizontal_Bounding_Boxes/*.xml
                Oriented_Bounding_Boxes/*.xml
        ImageSets/Main/
                test.txt  
                train.txt  
                val.txt
        JPEGImages/*.jpg

• To use DIOR datasets in object detection projects, the following can be run, to train CenterNet with DIOR dataset
• You can train on other networks by replacing the following Settings with
  • centernet_resnet101_dcnv2_140e_dior.py
  • centernet_resnet50_dcnv2_140e_dior.py
  • centernet_resnet18_dcnv2_140e_dior.py
    
  • faster_rcnn_r101_caffe_fpn_1x_dior.py
  • mask_rcnn_r50_caffe_fpn_mstrain-poly_1x_dior.py

cd mmdetection-2.24.1/
python tools/train.py configs/DIOR/centernet_resnet101_dcnv2_140e_dior.py

• Using DIOR dataset in fewshot learning, the following can be run, to train Meta-RCNN using DIOR query suport dataset

cd mmfewshot/
python tools/detection/train.py configs/detection/dior/meta-rcnn_r101_c4_8xb4_dior-split1_base-training.py      #base training
python tools/detection/train.py configs/detection/dior/meta-rcnn_r101_c4_8xb4_dior-split1_5shot-fine-tuning.py  #fine-tuning

• Training DIOR query suport dataset on the new implemented Meta-CenterNet can run the following

cd mmfewshot/
python tools/detection/train.py configs/detection/meta_centernet/dior/meta-centernet_r50_c4_8xb4_dior-split1_base-training.py       #base training
python tools/detection/train.py configs/detection/meta_centernet/dior/meta-centernet_r50_c4_8xb4_dior-split1_5shot-fine-tuning.py   #fine-tuning

• DIOR dataset

@article{li2020object,
  title={Object detection in optical remote sensing images: A survey and a new benchmark},
  author={Li, Ke and Wan, Gang and Cheng, Gong and Meng, Liqiu and Han, Junwei},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={159},
  pages={296--307},
  year={2020},
  publisher={Elsevier}
}

Pytorch

mmfewshot