Skip to content

Find logos in images and videos in just one-shot. Never be embarrassed again to say that you have a small data situation!

License

Notifications You must be signed in to change notification settings

Heldenkombinat/Logodetect

Repository files navigation

PyPI version DOI

Never be embarrassed again to say you have no big data! logodetect is a one-shot detection library to find logos of any kind in video and image data.

Here's a quick example of football footage that can detect all logos on jerseys and the sports field. Go check out our demo if you want to see it in action right away.

Introduction

There is plenty of literature on the use of deep-learning for detecting logos, so, additionally to sharing with you a couple of algorithms to get started with one-shot logo-detection, the aim of this project is to develop a flexible architecture to facilitate the comparison of different algorithms for one-shot object recognition.

The pipeline supports one or two stages. It is possible to only perform object-recognition, or to first perform object-detection and then object-recognition in a second stage.

The idea is that you can use a generic detector for a single class of objects (e.g. logos, traffic signs or faces) and then compare each of its detections with the exemplar, i.e., the sub-class that you are trying to recognize, to determine if both belong to the same sub-class (e.g. a concrete brand, a stop sign or the face of a loved one). To get started, we include two algorithms that you can play with. Both have a Faster-RCNN [1] in the first stage that performs object-detection and they differ in the second stage that performs object-recognition.

As a baseline, we bring the exemplars and the detections from the first stage to the same latent space (this reduces the course of dimensionality) and then simply measure the Euclidian or the cosine distance between both embeddings for object-recognition. Both inputs are considered to belong to the same sub-class if their distance is below a threshold determined in a preliminary analysis of the training dataset. The code also provides functionality to add various transformations, so you have the option to augment each exemplar with different transformations if you want. Simply add one or more exemplars into the data/exemplars folder that is generated after you've followed the installation instructions below, and you are good to go.

As a first reference against the baseline, we also provide a modified ResNet [2] for object-recognition that directly takes the exemplars and the detections from the first stage and predicts if both belong to the same sub-class. Similarly to [3], this network infers a distance metric after being trained with examples of different sub-classes, but instead of sharing the same weights and processing each input in a separate pass as in [4], it concatenates both inputs and processes them in one pass. This concept follows more closely the architecture proposed by [5], where the assumption is that the exemplars often have more high-frequency components than the detections, and therefore the model can increase its accuracy by learning a separate set of weights for each input. However, our proposed architecture splits the detection and classification tasks in two separate stages, which allows the use, and comparison, of different classifiers for the second stage.

The models that we are including in the repo achieved a reasonable performance after a few training epochs. However, if you would like to improve their performance you can find pointers to various datasets in [6], which can be used in the training part of this project.

References

[1] Ren et. al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (2016)
[2] He et. al. Deep Residual Learning for Image Recognition (2016)
[3] Hsieh et. al. One-Shot Object Detection with Co-Attention and Co-Excitation (2019)
[4] Koch et. al. Siamese Neural Networks for One-shot Image Recognition (2015)
[5] Bhunia et. al. A Deep One-Shot Network for Query-based Logo Retrieval (2019)
[6] Hoi et. al. LOGO-Net: Large-scale Deep Logo Detection and Brand Recognition with Deep Region-based Convolutional Networks (2015)

Installation

This library is intended for Linux-based OS, such as Ubuntu, and currently can't be run in Windows OS. It depends on libgl1, which you can install with apt-get install libgl1. Logodetect is available on PyPI, so you can simply run pip install logodetect to install it. Make sure you have Python version 3.7 or later installed and to have an up-to-date pip version by running pip install -U pip. Also, we recommend working with virtual environments, but that is ultimately up to you.

If you want to build logodetect from source, run

git clone git@github.com:Heldenkombinat/logodetect.git
cd logodetect
pip install -e ".[tests, dev]"

Depending on your system and setup, you might have to run the install command as sudo.

Usage

After successful installation, a CLI tool called logodetect becomes available to you. If you invoke logodetect without any arguments, you will get help on how to use it. To automatically download all models and data needed to test the application first run the following command in your clone of this repository:

export LOGOS_RECOGNITION=$(pwd)
logodetect init

which will download all files to the current working directory. Note that if you prefer another folder to download the data, please use the environment variable LOGOS_RECOGNITION accordingly. Consider putting this variable in your .bash_rc, .zshrc or an equivalent configuration file on your system. If you don't specify a folder, it will default to ~/.hkt/logodetect.

After running the logodetect init CLI, you'll find data and models relative to the specified folder in the following structure:

data/
    exemplars/
    exemplars_100x100/
    exemplars_100x100_aug/
    exemplars_hq/
    test_images/
    test_videos/
models/
    classifier_resnet18.pth
    detector.pth
    embedder.pth

If you're interested in training your own algorithms, it's a good idea to have a look at how the exemplar data is structured. For more on training, see the training folder and its readme.

The logodetect CLI tool comes with two main commands, namely video and image, both of which work fairly similarly. In each case you need to provide the input data for which you would like to detect logos, and the logo exemplars that you want to detect in the footage. To get you started, we've provided some demo data that you can use out of the box. That means you can simply run:

logodetect video

which should output the following text:

Rendering video: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 17/17 [00:00<00:00, 707.42it/s]
Moviepy - Building video /path/data/test_videos/test_video_small_50ms_output.mp4.
Moviepy - Writing video /path/data/test_videos/test_video_small_50ms_output.mp4

Moviepy - Done !
Moviepy - video ready /path/data/test_videos/test_video_small_50ms_output.mp4
All done! ✨ 🍰 ✨

to run one-shot detection on an example video, or you can run

logodetect image

to do so for an example image, which results in the following output:

Saved resulting image as /path/data/test_images/test_image_small_output.png.
All done! ✨ 🍰 ✨

If you want to use another video, you can do so with the -v option. Images can be provided with the -i option and custom exemplars are configured with the -e option. If you want to run logodetect with your own, custom configuration, please provide a JSON file (like the config.json in this repo) with the -c option. That means, if you want to run detection for custom video data with custom exemplars, you should use

logodetect video -v <path-to-video> -e <path-to-exemplars-folder> -c <path-to-custom-config-json>

Minimal web application for image recognition

To run a small web app locally in your browser to upload images to recognize, simply run

python app.py

and navigate to https://localhost:5000 in the browser of your choice. Also, we've hosted an online demo for you here.

On top of that, the aws folder explains in detail how to host this application yourself on Amazon Web Services. This minimalistic application can of course be extended to your own needs at any point.

Full set of CLI commands and help pages

In the last section we have already discussed the three commands exposed to users through the logodetect CLI tool, namely init, image, and video. While init does not take any parameters, the other two need a bit more explanation. Below you find the complete API reference from the respective help pages of our CLI.

Images

logodetect image --help
Usage: logodetect image [OPTIONS]

Options:
  -i, --image_filename TEXT   path to your input image
  -c, --config_file TEXT      path to file containing a logodetect config JSON
  -o, --output_appendix TEXT  string appended to your resulting file
  -e, --exemplars TEXT        path to your exemplars folder
  --help                      Show this message and exit.

Videos

logodetect video --help
Usage: logodetect video [OPTIONS]

Options:
  -v, --video_filename TEXT   path to your input video
  -c, --config_file TEXT      path to file containing a logodetect config JSON
  -o, --output_appendix TEXT  string appended to your resulting file
  -e, --exemplars TEXT        path to your exemplars folder
  --help                      Show this message and exit.

Core abstractions

logodetect works with a two-phased approach. In the first phase, objects get detected with a Detector, and then get compared to and identified with exemplars in a Classifier. Both phases get integrated into the inference pipeline using a single Recognizer, in which we detect potential overlay boxes in video frames or images and then the detected boxes get labeled according to their classification.

Configuration

The specific parameter settings of the algorithms used in logodetect, i.e. options for all of our detectors, classifiers, data augmenters, and system devices used, can be changed by providing a config.json file with the -c flag in the main CLI commands explained above. The example config.json file explains the options you have and what exactly you can modify in logodetect.

Notebooks

You can find exemplary jupyter notebooks from the logodetect project in the notebooks/ folder. If you're interested in training new models, then training/notebooks/ might interest you.

Docker support

If you prefer to work with Docker, build an image and run it like this:

docker build . -t logodetect
docker run -e LOGOS_RECOGNITION=/app -p 5000:5000 -t logodetect

Important: this assumes that you have previously downloaded all data and models right next to the Dockerfile in the local copy of this repo.

Automatic code linting with black

This project uses black for code linting. To install the git pre-commit hook for black, simply run

pre-commit install

from the base of this repository. This will run (and fail in case of grave errors) black each time you make a commit. Once CI is up for this project, we will ensure this hook runs on each CI pass. To manually use black on a file, use black <path-to-file>.

Running tests

Run all tests with pytest, or just run the quicker unit test suite with

pytest -m unit

or all longer-running integration tests with

pytest -m integration

Building the paper locally

docker run --rm \
    --volume $PWD:/data \
    --user $(id -u):$(id -g) \
    --env JOURNAL=joss \
    openjournals/paperdraft

Support

For support, issues and contributions, please follow the guidelines in SUPPORT.md.