Faster-RCNN

Implementation of MultiScaleRoIAlign and MultiScaleRoIAlign in Faster RCNN, code mainly comes from deep-learning-for-image-processing/pytorch_object_detection/faster_rcnn. Specific details analysis about the implementation of MultiScaleRoIAlign and MultiScaleRoIAlign, please refer to the blog.

Prepare environment：

• Python3.6/3.7/3.8
• Pytorch1.7.1
• pycocotools
• Ubuntu or CentOS
• detailed environment config refer to requirements.txt

File structure：

  ├── backbone: backbone files
  ├── network_files: Faster R-CNN framework files
  ├── train_utils: training and validation modules
  ├── my_dataset.py: custom dataset for reading VOC dataset
  ├── train_mobilenet.py: using MobileNetV2 as backbone
  ├── train_resnet50_fpn.py: using resnet50+FPN as backbone
  ├── train_multi_GPU.py: using multiGPU for training
  ├── predict.py: prediction script
  ├── validation.py: validation script and generated record_mAP.txt
  └── pascal_voc_classes.json: pascal_voc label json

Download pre-trained weights（place it in the backbone folder）：

• MobileNetV2 weights download link: https://download.pytorch.org/models/mobilenet_v2-b0353104.pth
• Resnet50 weights download link: https://download.pytorch.org/models/resnet50-0676ba61.pth
• ResNet50+FPN weights download link: https://download.pytorch.org/models/fasterrcnn_resnet50_fpn_coco-258fb6c6.pth

Note that the pretraining weights downloaded should be renamed, for example, the downloaded weight is named fasterrcnn_resnet50_fpn_coco-258fb6c6.pth and should be renamed fasterrcnn_resnet50_fpn_coco.pth.

prepare datasets

The dataset used in this repo is Pascal VOC2012 train/val. Dataset download link at here.

training process

• ensure that the dataset is ready

• ensure that the pretraining weights are downloaded properly

• To using mobilenetv2 as backbone, directly use the train_mobilenet.py

• To using resnet50_fpn as backbone, directly use the train_res50_fpn.py

• If you want to use multiGPU training, use the following command

  CUDA_VISIBLE_DEVICES=0,3 torchrun --nproc_per_node=2 train_multi_GPU.py --data-path [PASCAL_VOC_2012 datasets path] -b [batch_size]