Kaggle IEEE's Signal Processing Society - Camera Model Identification

Implementation of camera model identification system by team "[ods.ai] GPU_muscles" (2nd place overall in Kaggle competition IEEE's Signal Processing Society - Camera Model Identification and 1st place among student eligible teams).

Should any questions arise regarding the solution, please do not hesitate to contact me on Telegram or via e-mail [email protected]

Our team

• Artur Kuzin [linkedin]
• Valeriy Babushkin [linkedin]
• Artur Fattakhov [kaggle]
• Ilya Kibardin [linkedin]
• Andrey Kiselev [linkedin]

Requirements

To train models and get predictions the following is required:

• OS: Ubuntu 16.04
• Python 3.6
• Hardware:
  • Any decent modern computer with x86-64 CPU,
  • 32 GB RAM
  • 4 x Nvidia GeForce GTX 1080 Ti

Installation

1. Install required OS and Python
2. Install packages with pip install -r requirements.txt
3. Create data folder at the root of the repository. Place train dataset from Kaggle competition to data/train. Place test dataset from Kaggle competition to data/test. Place additional validation images to data/val_images.
4. Place se_resnet50.pth and se_resnext50.pth to imagenet_pretrain folder.
5. Place the following final weights to final_weights folder:
   • densenet161_28_0.08377413648371115.pth
   • densenet161_55_0.08159203971706519.pth
   • densenet161_45_0.0813179751742137.pth
   • dpn92_tune_11_0.1398952918197271.pth
   • dpn92_tune_23_0.12260739478774665.pth
   • dpn92_tune_29_0.14363511492280367.pth

Producing the final submission

Run bash final_submit.sh -d <folder with test images> -o <output .csv filename>

Training ensemble from scratch

This section describes the steps required to train our ensemble.

1. Download external dataset

Images from both Yandex.Fotki and Flickr are essential for reproducing our solution.

Downloading images from Yandex.Fotki

Run bash download_from_yandex.sh

Downloading images from Flickr

Unfortunately, this step involves some manual actions.

1. cd into downloader/flickr
2. For every model go to the telephone model group page from flickr_groups.txt. Scroll every gallery page to the end and download as html file to the corresponding folder. As a result you will have a set of folders with .html files corresponding to a specific phone model at html_pages folder.
3. Run python pages_to_image_links.py. The result of the script will be folder links of .csv files with links to photos of each phone model.
4. Run python download_from_links.py to download images from the links received in the previous paragraph (previous two steps could be skipped, because the links folder already contains necessary files).

2. Filter external dataset

Run bash filter.sh

3. Train the ensemble

1. Download and filter external dataset as described above.
2. Run bash init_train.sh to train 9 models.
3. Run bash make_pseudo.sh to get predictions from these models for images at data/test and create pseudo labels.
4. Run bash final_train.sh to train the same 9 models but using pseudo labels this time.
5. Run bash predict.sh -d <folder with test images> -o <output .csv filename> to get predictions from the ensemble.