TECH ARENA POLAND 2023

Huawei Tech Arena is an annual competition across Poland universities.

As an industry leader in information, communications technology, infrastructure and smart devices, Huawei is eager to work with the next generation of innovation thinkers.

We are happy to present a 6-week online challenge with 1 objective:

Multimodal File Search Challenge

REGISTRATIONS CLOSED

MORE ABOUT THE CHALLENGE

The term "multimodal file search" refers to a file search process that utilizes various modalities, such as text, image, sound, and video, to deliver the most accurate results. It works by combining and analyzing different types of input data, allowing for a more complex and comprehensive understanding of the search context.

For example, a multimodal file search might allow a user to search documents not only based on textual metadata, such as file name or tags, but also based on the content of those files, such as text contained within a document, graphical elements in a PDF file, or even dialogues from an audio or video recording.

This can significantly facilitate finding specific files in large data sets, especially in those that contain various types of files.

In this contest, we will focus on query-based multimodal retrieval task by using the combinations of image and text model.

Given a collection of multimodal data (pairs of image and text descriptions of the image) and search query in natural language, the task is to design and implement a model to discover top images in this collection that are relevant to the query.

Images most relevant to the query are regarded as the ground truth of the query, which are expected to be top-ranked by the participating models. The applications implementing the model should run on a standalone desktop or laptop.

Challenge Phases

There will be 2 phases during the competition – The Training Phase & The Final Competition Phase.

Phase 1 - Training Phase

Phase 2 - Competition Phase

The datasets for both phases will be provided via the competition webpage.

The datasets consist of the Training Set, a Validation Set and a Testing Set. The Training Set and Validation Set will be released in the training phase, and the Testing Set will be revealed in the final phase.

Training Set

The Training Set consists of pairs of query and ground-truth image text descriptions. These sample pairs can be taken as the positive examples to train your retrieval model.

Validation Set

The Validation Set consists of a number of queries and the Testing Set consist of another queries. For each query in these sets, there is a candidate pool of images with paired text description, and the ground-truth results for each evaluation query.

Testing Set

The Testing Set is the same format to the Validation Set, but it will only been revealed in the final competition phase and the ground-truth results for the Testing Set will not be released.

REGISTRATIONS CLOSED

Deliverables and Judging Criteria

Submit Format

Evaluation