RetinaNet-PyTorch - A RetinaNet Pytorch Implementation on remote sensing images and has the similar mAP result with RetinaNet in MMdetection

Last update: Nov 19, 2022

Overview

🚀 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 Inshore image.

Detection Performance on Offshore image.

🎯 Experiment

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 data set BaiduYun extraction code=vqaw (ground-truth)
gt labels for train data set BaiduYun extraction code=datk (train-ground-truth)
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

Installation

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
make

Demo

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  
$ROOT_PATH
    -SSDD_data/
        -train/  # train set 
	    -*.jpg
	-val/  # val set
	    -*.jpg
	-annotations/  # gt label in json format (for coco evaluation method)  
	    -instances_train.json  
	    -instances_val.json  
	-ground-truth/  
	    -*.txt  # gt label in txt format (for voc evaluation method and evaluae inshore and offshore scence)  
	-train-ground-truth/
	    -*.txt  # gt label in txt format (for voc evaluation method)
    -SSDD_data_InShore/  
        -images/
	    -*.jpg  # inshore scence images
	-ground-truth/
	    -*.txt  # inshore scence gt labels  
    -SSDD_data_OffShore/  
        -images/  
	    -*.jpg  # offshore scence images
	-ground-truth/  
	    -*.txt  # offshore scence gt labels

    -Output/
        -checkpoints/
	    - the path of saving tensorboard log event
	-evaluate/  
	    - 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

Train

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.

Evaluation

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

💡 Inferences

Thanks for these great work.
https://github.com/ming71/DAL
https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch

RetinaNet-PyTorch - A RetinaNet Pytorch Implementation on remote sensing images and has the similar mAP result with RetinaNet in MMdetection

Related tags

Overview

🚀 RetinaNet Horizontal Detector Based PyTorch

🌟 Performance of the implemented RetinaNet Detector

Detection Performance on Inshore image.

Detection Performance on Offshore image.

🎯 Experiment

💥 Get Started

Installation

A. Install requirements:

B. Install nms module:

Demo

A. Set project's data path

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

Train

A. Prepare dataset

B. Manual set project's hyper parameters

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

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

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

Evaluation

A. evaluate model performance on val set.

B. evaluate model performance on InShore and Offshore sences.

C. evaluate model FPS

💡 Inferences

Owner

Fang Zhonghao

ZeroGen: Efficient Zero-shot Learning via Dataset Generation

A quick recipe to learn all about Transformers

GULAG: GUessing LAnGuages with neural networks

Python scripts form performing stereo depth estimation using the HITNET model in ONNX.

A supplementary code for Editable Neural Networks, an ICLR 2020 submission.

PyTorch implementation of ICLR 2022 paper PiCO: Contrastive Label Disambiguation for Partial Label Learning

This provides the R code and data to replicate results in "The USS Trustee’s risky strategy"

Neural implicit reconstruction experiments for the Vector Neuron paper

[Pedestron] Generalizable Pedestrian Detection: The Elephant In The Room. @ CVPR2021

Code for paper: Group-CAM: Group Score-Weighted Visual Explanations for Deep Convolutional Networks

Keras-tensorflow implementation of Fully Convolutional Networks for Semantic Segmentation（Unfinished）

Use deep learning, genetic programming and other methods to predict stock and market movements

Video-Music Transformer

Implementation of Vision Transformer, a simple way to achieve SOTA in vision classification with only a single transformer encoder, in Pytorch

Creating Multi Task Models With Keras

Detection of PCBA defect

PyTorch3D is FAIR's library of reusable components for deep learning with 3D data

To model the probability of a soccer coach leave his/her team during Campeonato Brasileiro for 10 chosen teams and considering years 2018, 2019 and 2020.

A code repository associated with the paper A Benchmark for Rough Sketch Cleanup by Chuan Yan, David Vanderhaeghe, and Yotam Gingold from SIGGRAPH Asia 2020.

A Large-Scale Dataset for Spinal Vertebrae Segmentation in Computed Tomography