This network is to be trained to detect and localize the ships from the satellite image. Dataset are obtained from Airbus Ship Detection Challenge with segmentation label.
For the phase 1 of the project, we train a VGG with normlization and resnet to classify a image to has ship and no ship to classes. This phase help us to evaluate the learning capability of the base network.
In the phase 2, base on the base model we buid our own model on top of a faster-rcnn project from a faster-rcnn project at https://github.com/jwyang/faster-rcnn.pytorch.git. to localize the ships
Host computer: Ubuntu 18.04 CUDA 10.0 in host
Docker image: nvidia/cuda:9.0-cudnn7-devel Docker images: CUDA 9.0 python 3.6
A project %ROOT folder should contains at least dataset, models and src folders
Data can be downloaded from Airbus Ship Detection Challenge. The downloaded data then renamed to dataset under $ROOT
run $ROOT/src/data_loader/gt_box_gen.py to convert segmentation mask to bounding box for training.
Go to %ROOT/src/base, you can either open the base python notebook or run the following on terminal
Training: python base.py -t Evaluate: python base.py -e Demo: python base.py -d
For just 1 epoch The VGG model can achieve 91.5% accuracy The Resnet model can achieve 92.2% accuracy
Because VGG comsume lesser memory, and perform almost the same, we choose VGG as our base in the next phase
We used five pretrained models, each of them are saved as 1 to 5 epoch during training
Download models folder from https://drive.google.com/drive/folders/1X-bitFl5E6uOfCE3jnhE3XC5XvIQJbae?usp=sharing and put them into the $ROOT/
As pointed out by ruotianluo/pytorch-faster-rcnn, choose the right -arch in make.sh file, to compile the cuda code:
| GPU model | Architecture |
|---|---|
| TitanX (Maxwell/Pascal) | sm_52 |
| GTX 960M | sm_50 |
| GTX 1080 (Ti) | sm_61 |
| Grid K520 (AWS g2.2xlarge) | sm_30 |
| Tesla K80 (AWS p2.xlarge) | sm_37 |
More details about setting the architecture can be found here or here
Install all the python dependencies using pip:
pip install -r requirements.txt
Compile the cuda dependencies using following simple commands:
cd lib
sh make.sh
It will compile all the modules you need, including NMS, ROI_Pooing, ROI_Align and ROI_Crop. The default version is compiled with Python 2.7, please compile by yourself if you are using a different python version.
This code has to be run with a computer has Nvidia GPU Navigate to src folder
To train a faster R-CNN model with vgg16 on kaggle ship dataset, simply run: python trainval_net.py --cuda
If you want to evlauate the detection recall and precision of a pre-trained vgg16 model on ships dataset test set, simply run
python test_net.py --cuda
The test data are randomly selected from test set and saved to %ROOT/demo/output after processed
If you want to visualize detected output simply run python test_net.py --cuda --vis