?RetinaNet Horizontal Detector Based PyTorch

This is a horizontal detector RetinaNet implementation on remote sensing ship dataset (SSDD).
This re-implemented retinanet has the almost the same mAP(iou=0.25, score_iou=0.15) with the MMdetection.
RetinaNet Detector original paper link is here.

?Performance of the implemented RetinaNet Detector

Detection Performance on Offshore image.

Detection Performance on Inshore image.


The SSDD dataset, well-trained retinanet detector, resnet-50 pretrained model on ImageNet, loss curve, evaluation metrics results are below, you could follow my experiment.

  • SSDD dataset BaiduYun extraction code=pa8j
  • gt labels for eval BaiduYun extraction code=vqaw
  • well-trained retinanet detector weight file BaiduYun extraction code=b0e1
  • pre-trained ImageNet resnet-50 weight file BaiduYun extraction code=mmql
  • evaluation metrics(iou=0.25, score_iou=0.15)
Batch Size Input Size mAP (Mine) mAP (MMdet) Model Parameters
32 416 x 416 0.8828 0.8891 32.2 M
  • Other metrics (Precision/Recall/F1 score)
Precision (Mine) Precision (MMDet) Recall (Mine) Recall (MMdet) F1 score (Mine) F1 score(MMdet)
0.8077 0.8502 0.9062 0.91558 0.8541 0.8817
  • loss curve

  • mAP metrics on training set and val set

  • learning rate curve (using warmup lr rate)

?Get Started


A. Install requirements:

conda create -n retinanet python=3.7
conda activate retinanet
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch
pip install -r requirements.txt  

Note: If you meet some troubles about installing environment, you can see the check.txt for more details.

B. Install nms module:

cd utils/HBB_NMS_GPU


A. Set project’s data path

you should set project’s data path in config.py first.

# config.py
# Note: all the path should be absolute path.  
data_path = r'/$ROOT_PATH/SSDD_data/'  # absolute data root path  
output_path = r'/$ROOT_PATH/Output/'  # absolute model output path  
inshore_data_path = r'/$ROOT_PATH/SSDD_data_InShore/'  # absolute Inshore data path  
offshore_data_path = r'/$ROOT_PATH/SSDD_data_OffShore/'  # absolute Offshore data path  

# An example  
        -train/  # train set 
	-val/  # val set
	-annotations/  # gt label in json format (for coco evaluation method)  
	    -*.txt  # gt label in txt format (for voc evaluation method and evaluae inshore and offshore scence)  
	    -*.jpg  # inshore scence images
	    -*.txt  # inshore scence gt labels  
	    -*.jpg  # offshore scence images
	    -*.txt  # offshore scence gt labels

	    - the path of saving tensorboard log event
	    - the path of saving model detection results for evaluate (coco/voc/inshore/offshore)  

B. you should download the well-trained SSDD Dataset weight file.

# download and put the well-trained pth file in checkpoints/ folder 
# and run the simple inferene script to get detection result  
# you can find the model output predict.jpg in show_result/ folder.  

python show.py --chkpt 54_1595.pth --result_path show_result --pic_name demo1.jpg  


A. Prepare dataset

you should structure your dataset files as shown above.

B. Manual set project’s hyper parameters

you should manual set projcet’s hyper parameters in config.py

1. data file structure (Must Be Set !)  
   has shown above.  

2. Other settings (Optional)  
   if you want to follow my experiment, dont't change anything.  

C. Train RetinaNet detector on SSDD dataset with pretrianed resnet-50 from scratch

C.1 Download the pre-trained resnet-50 pth file

you should download the pre-trained ImageNet Dataset resnet-50 pth file first and put this pth file in resnet_pretrained_pth/ folder.

C.2 Train RetinaNet Detector on SSDD Dataset with pre-trained pth file

# with batchsize 32 and using voc evaluation method during training for 50 epochs  
python train.py --batch_size 32 --epoch 50 --eval_method voc  
# with batchsize 32 and using coco evalutation method during training for 50 epochs  
python train.py --batch_size 32 --epoch 50 --eval_method coco  

Note: If you find classification loss change slowly, please be patient, it's not a mistake.


A. evaluate model performance on val set.

python eval.py --device 0 --evaluate True --FPS False --Offshore False --Inshore False --chkpt 54_1595.pth

B. evaluate model performance on InShore and Offshore sences.

python eval.py --device 0 --evaluate False --FPS False --Offshore True --Inshore True --chkpt 54_1595.pth

C. evaluate model FPS

python eval.py --device 0 --evaluate False --FPS True --Offshore False --Inshore Fasle --chkpt 54_1595.pth


Thanks for these great work.


View Github