From text to image and back in Seznam.cz

Image-Text Alignment: Learning Representations for Image Search

Abstract:
Image search engine is a multimodal environment that connects the user’s text input (queries) with images. Aligned vector representations of both modalities significantly help with image retrieval from a large index and as a signal for relevance ranking model. We will discuss our approach of using state-of-the-art vision-language representation learning techniques to create a model suitable for deployment as a part of Seznam’s image search. Tackling obstacles such as Czech language, raw text data and hardware limitations along the way.

About Speaker:
Martin Habrovec is a senior researcher at Seznam.cz where he has worked on a diverse set of successfully deployed projects in NLP, computer vision and information retrieval. His previous experiences also include big data processing and analytics.

Traffic signs detection from panoramic images for Mapy.cz

Abstract:
Panoramic images are a popular feature of Mapy.cz. With the increase of their quality and coverage they can also be used as a source to extract additional information using computer vision. One piece of such information is the position, type and text of traffic signs which can be further leveraged in navigation systems. In this talk we will introduce a real-life application of object detection in a specific domain of 360 degree panoramic images and the practical challenges we faced along the way.

About Speaker:
Dalibor Skácel is a senior researcher at Seznam.cz where he works on specialized machine learning solutions for Mapy.cz including text and image processing, search results ranking and destination recommendation engines.