A5002095484- Advanced Image and Video Retrieval Techniques
- Image Retrieval and Classification Techniques
- Multimodal Machine Learning Applications
- Domain Adaptation and Few-Shot Learning
- Video Analysis and Summarization
- Data Management and Algorithms
- Medical Image Segmentation Techniques
- Visual Attention and Saliency Detection
- Web Data Mining and Analysis
- Face recognition and analysis
- Distributed and Parallel Computing Systems
- Advanced Neural Network Applications
- Optimization and Search Problems
- Video Surveillance and Tracking Methods
- Face and Expression Recognition
- Algorithms and Data Compression
- Robotics and Sensor-Based Localization
- Scientific Computing and Data Management
HTW Berlin - University of Applied Sciences
2019-2024
Istituto di Scienza e Tecnologie dell'Informazione "Alessandro Faedo"
2023
Joanneum Research
2023
Charles University
2023
University of Klagenfurt
2023
As datasets and the dimensionality of feature vectors continue to grow, Approximate Nearest Neighbor Search (ANNS) in large multimedia databases becomes increasingly relevant. Graph-based approaches have demonstrated offer best trade-off between retrieval precision search time. Despite their ability deliver times several orders magnitude faster than exact techniques, existing methods suffer from slow constructions speeds or high memory requirements. This paper presents a continuous refining...
Abstract CLIP-based text-to-image retrieval has proven to be very effective at the interactive video competition Video Browser Showdown 2022, where all three top-scoring teams had implemented a variant of CLIP model in their system. Since performance these systems was quite close, this post-evaluation designed get better insights on differences and compare text-query engines by introducing slight modifications original settings. An extended analysis overall results systems’ functionalities...
Building on our success with the Vibro video search system in Video Browser Showdown, we are beginning a new effort by applying technologies to Lifelog Search Challenge for first time. Our approach is treat lifelog data collected given day as frames of continuous clip. While have essentially adopted text-to-image and image-to-image from Vibro, introduced various metadata filters complement capabilities. goal increase efficiency image searches within dataset integrating these improvements....
Contrastive Language and Image Pairing (CLIP), a transformative method in multimedia retrieval, typically trains two neural networks concurrently to generate joint embeddings for text image pairs. However, when applied directly, these models often struggle differentiate between visually distinct images that have similar captions, resulting suboptimal performance image-based similarity searches. This paper addresses the challenge of optimizing CLIP various search scenarios, while maintaining...
Recent advances in computer vision research led to large foundation models that generalize a broad range of image domains and perform exceptionally well various based tasks. However, content-based image-to-image retrieval is often overlooked this context. This paper investigates the effectiveness different on two challenging nearest neighbor search-based tasks: zero-shot k-NN classification. A benchmark for evaluating performance encoders their pre-training methods established, where...
One of the key challenges deep learning based image retrieval remains in aggregating convolutional activations into one highly representative feature vector. Ideally, this descriptor should encode semantic, spatial and low level information. Even though off-the-shelf pre-trained neural networks can already produce good representations combination with aggregation methods, appropriate fine tuning for task has shown to significantly boost performance. In paper we present a simple yet effective...
Nowadays stock photo agencies often have millions of images. Non-stop viewing 20 million images at a speed 10 per second would take more than three weeks. This demonstrates the impossibility to inspect all and difficulty get an overview entire collection. Although there has been lot effort improve visual image search, is little research support for exploration. Typically, users start "exploring" collection with keyword search or example similarity search. Both searches lead long unstructured...
Due to the size of today's image collections it can be challenging fully understand their content. Recent technological advances have enabled efficient visual search. These systems use joint and textual feature vectors identify similar images based on queries or text descriptions. Despite effectiveness, high-dimensional lead long search times for large collections. In this demonstration, we propose a solution that significantly reduces increases efficiency system. By combining two separate...
For approximate nearest neighbor search, graph-based algorithms have shown to offer the best trade-off between accuracy and search time. We propose Dynamic Exploration Graph (DEG) which significantly outperforms existing in terms of exploration efficiency by combining two new ideas: First, a single undirected even regular graph is incrementally built partially replacing edges integrate vertices update old neighborhoods at same Secondly, an edge optimization algorithm used continuously...
The native macOS application PicArrange integrates state-of-the-art image sorting and similarity search to enable users get a better overview of their images. Many file management features have been added make it tool that addresses full workflow. A modification the Self Sorting Map algorithm enables list-like arrangement without loosing visual sorting. Efficient calculation storage as well use many APIs result in an is fluid use.
Visually sorted grid layouts provide an efficient method for organizing high-dimensional vectors in two-dimensional space by aligning spatial proximity with similarity relationships. This approach facilitates the effective sorting of diverse elements ranging from data points to images, and enables simultaneous visualization a significant number elements. However, on grids is challenge due its high complexity. Even small 8-by-8 64 elements, possible arrangements exceeds 1.3 * 10^89 - more...
Abstract CLIP-based text-to-image retrieval has proven to be very effective at the interactive video competition Video Browser Showdown 2022, where all three top-scoring teams had implemented a variant of CLIP model in their system. Since performance these systems was quite close, this post-evaluation designed get better insights on differences and compare text-query engines by introducing slight modifications original settings. An extended analysis overall results systems' functionalities...
Image collections today often consist of millions images, making it impossible to get an overview the entire content. In recent years, we have presented several demonstrators for graph-based systems allowing image search and a visual exploration collection. Meanwhile, very powerful also joint visual-textual feature vectors been developed, which are suitable finding similar images query or according textual description. A drawback these is that they high number dimensions, leads long times,...
In recent years, deep metric learning has achieved promising results in high dimensional semantic feature embeddings where the spatial relationships of vectors match visual similarities images. Similarity search for images is performed by determining with smallest distances to a query vector. However, retrieval quality does not depend on actual vectors, but rather ranking order from similar this paper, we introduce algorithm that focuses identifying and modifying those most strongly affect...