Ships Classification using PyTorch

This project was carried out for the Leonardo Labs Kaggle Competition and was able to achieve the 2nd place with a private score of 0.98102. The goal was to classify images of ships according to their type (7 different classes).

Approach

The strategy chosen was to train different neural network architectures and then obtain a single prediction by majority voting. Other strategies that have been adopted include training an ensemble of neural networks after combining the outputs and the use of Visual Transformers.

Dataset

In the original training set there are images out of context that must be removed, this process has already been done by us, in the following csv file only the files and the respective categories of images concerning ships are listed:

!gdown 1AobspqrM3TYKw3DvWUpsH3GHhCCMMovg

The dataset can be downloaded with the following command:

!gdown 1hukMWTFj2aSqx2jBh42R-Y6UXrSw60Nj

The preprocess of the image is done image by image by us for a fine-tuned cleaning and a first automatic cleaning using VGG16 to extract the feature of the image and use a unsupervised method to cluster the image in different class.

Here are some examples:

battleships	coast-guard	containerships	cruise-ships	drilling-rigs	motor-yachts	submarines

Models

Models	Paper	Link
ViT B 16	https://arxiv.org/abs/2010.11929	https://github.com/lukemelas/PyTorch-Pretrained-ViT
Resnet152	https://arxiv.org/abs/1512.03385	https://pytorch.org/vision/0.12/generated/torchvision.models.resnet152.html
ConvNeXt	https://arxiv.org/abs/2201.03545	https://pytorch.org/vision/stable/models/convnext.html
ResNeXt	https://arxiv.org/abs/1611.05431	https://pytorch.org/hub/pytorch_vision_resnext/
SE-ResNeXt	https://arxiv.org/abs/1709.01507v4	https://rwightman.github.io/pytorch-image-models/models/seresnext/
Xception	https://arxiv.org/abs/1610.02357	https://rwightman.github.io/pytorch-image-models/models/xception/

Team members

• Giulio D'Erasmo
• Arturo Ghinassi
• Andrea Potì
• Amedeo Ranaldi
• Camilla Savarese