SOMHunter opensource

The latest version of the SOMHunter tool is now open-source and available here under the GPLv2 license.

This is an open-source version of the SOMHunter video search and retrieval engine, slightly simplified from the prototype that was featured at Video Browser Showdown 2020 in Daejeon, Korea (see https://videobrowsershowdown.org/ ).

Main features:

• a very simple demonstration dataset is packed in the repository, indexed V3C1 dataset will be available for download
• keyword search, similarity-based search
• several ways of browsing and re-scoring the results (including SOMs)
• VBS-compatible submission and logging API clients

SOMHunter is licensed as free software, under the GPLv2 license. This grants you freedom to use the software for many research purposes and publish the results. For exploring and referencing the original work, you may find some of the following articles helpful:

• Kratochvíl, M., Veselý, P., Mejzlík, F., & Lokoč, J. (2020, January). SOM-Hunter: Video Browsing with Relevance-to-SOM Feedback Loop. In International Conference on Multimedia Modeling (pp. 790-795). Springer, Cham.
• Mejzlík, F., Veselý, P., Kratochvíl, M., Souček, T., & Lokoč, J. (2020, June). SOMHunter for Lifelog Search. In Proceedings of the Third Annual Workshop on Lifelog Search Challenge (pp. 73-75).
• Kratochvil, M., Mejzlík, F., Veselý, P., Soućek, T., & Lokoć, J. (2020, October). SOMHunter: Lightweight Video Search System with SOM-Guided Relevance Feedback. In Proceedings of the 28th ACM International Conference on Multimedia (pp. 4481-4484).
• Veselý, P., Mejzlík, F., & Lokoč, J. (2021, June). SOMHunter V2 at Video Browser Showdown 2021. In International Conference on Multimedia Modeling (pp. 461-466). Springer, Cham.

Try SOMHunter from Docker

You can get a working SOMHunter copy from Docker:

docker pull exaexa/somhunter:v0.1
docker run -ti --rm -p8080:8080 exaexa/somhunter:v0.1

After that, open your browser at http://localhost:8080, and use login som and password hunter.

Installation from source

Prerequisites:

• a working installation of Node.js with some variant of package manager (either npm or yarn)
• Python 3
• C++ compiler
• libcurl (see below for installation on various platforms)

After cloning this repository, change to the repository directory and run

npm install
npm run start

(Optionally replace npm with yarn.)

If everything goes all right, you can start browsing at http://localhost:8080/ . The site is password-protected by default, you can use the default login som and password hunter, or set a different login in config/user.js.

Getting the dependencies on UNIX systems

You should be able to install all dependencies from the package management. On Debian-based systems (including Ubuntu and Mint) the following should work:

apt-get install build-essential libcurl4-openssl-dev nodejs yarnpkg

The build system uses pkg-config to find libCURL -- if that fails, either install the CURL pkgconfig file manually, or customize the build configuration in core/binding.gyp to fit your setup.

Similar (similarly named) packages should be available on most other distributions.

Getting the dependencies on Windows

The build systems expects libCURL to reside in c:\Program Files\curl\. You may want to install it using vcpkg as follows:

• download and install vcpkg
• install and export libCURL:

vcpkg install curl:x64-windows
vcpkg export --raw curl:x64-windows

• copy the directory with the exported libCURL to c:\Program Files\.

Alternatively, you can use any working development installation of libCURL by filling the appropriate paths in core/binding.gyp.

Build problems

We have tested SOMHunter on Windows and several different Linux distributions, which should cover a majority of target environments. Please report any errors you encounter using the GitHub issue tracker, so that we can fix them (and improve the portability of SOMHunter).

Building the Docker image

The installation is described in Dockerfile; you should be able to get a working, correctly tagged (and possibly customized) image by running this in your directory:

docker build -t somhunter:$(git describe --always --tags --dirty=-$USER-$(date +%Y%m%d-%H%M%S)) .

Customizing SOMHunter

The program is structured as follows:

• The frontend requests are routed in app.js to views and actions in routes/somhunter.js, display-specific routes are present in routes/endpoints.js
• The views (for the browser) are rendered in views/somhunter.ejs
• Node.js "frontend" communicates with C++ "backend" that handles the main data operations; the backend source code is in core/; the main API is in core/SomHunterNapi.h (and .cpp)
• The backend implementation is present in core/src/ which contains the following modules (.cpp and .h):
  • SomHunter -- main data-holding structure with the C++ version of the wrapper API
  • Submitter -- VBS API client for submitting search results for the competition, also contains the logging functionality
  • DatasetFrames -- loading of the dataset description (frame IDs, shot IDs, video IDs, ...)
  • DatasetFeatures -- loading of the dataset feature matrix
  • KeywordRanker -- loading and application of W2VV keywords (see Li, X., Xu, C., Yang, G., Chen, Z., & Dong, J. (2019, October). W2VV++ Fully Deep Learning for Ad-hoc Video Search. In Proceedings of the 27th ACM International Conference on Multimedia (pp. 1786-1794).)
  • RelevanceScores -- maintenance of the per-frame scores and feedback-based re-ranking
  • SOM and AsyncSom -- SOM implementation, background worker that computes the SOM

Additional minor utilities include:

• config.h that contains various #defined constants, including file paths
• log.h which defines a relatively user-friendly logging with debug levels
• use_intrins.h and distfs.h define fast SSE-accelerated computation of vector-vector operations (provides around 4x speedup for almost all computation-heavy operations)
• main.cpp, which is not compiled-in by default, but demonstrates how to run the SOMHunter core as a standalone C++ application.

HOW-TOs

• Adding a new display type
• Modifying the re-scoring functionality

Datasets

The repository contains a (very small) pre-extracted indexed dataset (see https://doi.org/10.1109/ICMEW.2015.7169765 for dataset details). That should be ready to use.

We can provide a larger pre-indexed dataset based on the V3C1 video collection, but do not provide a direct download due to licensing issues. Please contact us to get a downloadable link. You will need to have the TRECVID data use agreement signed; see https://www-nlpir.nist.gov/projects/tv2019/data.html#licenses for details.

Using custom video data

You may set up the locations of the dataset files in config.json. The thumbnails of the extracted video frames must be placed in directory public/thumbs/, so that they are accessible from the browser. (You may want to use a symbolic link that points to the thumbnails elsewhere, in order to save disk space and IO bandwidth.)

Description of extracting data from custom dataset is available in directory extractor/ with a separate README.