Giorgos Kordopatis-Zilos

This work introduces ILIAS, a new test dataset for Instance-Level Image retrieval At Scale. It is designed to evaluate the ability of current and future foundation models and retrieval techniques to recognize particular objects. The key benefits over existing datasets include large scale, domain diversity, accurate ground truth, and a performance that is far from saturated. ILIAS includes query and positive images for 1,000 object instances, manually collected to capture challenging conditions and diverse domains. Large-scale retrieval is conducted against 100 million distractor images from YFCC100M. To avoid false negatives without extra annotation effort, we include only query objects confirmed to have emerged after 2014, i.e. the compilation date of YFCC100M. An extensive benchmarking is performed with the following observations: i) models fine-tuned on specific domains, such as landmarks or products, excel in that domain but fail on ILIAS ii) learning a linear adaptation layer using multi-domain class supervision results in performance improvements, especially for vision-language models iii) local descriptors in retrieval re-ranking are still a key ingredient, especially in the presence of severe background clutter iv) the text-to-image performance of the vision-language foundation models is surprisingly close to the corresponding image-to-image case.

This work investigates the problem of instance-level image retrieval re-ranking with the constraint of memory efficiency, ultimately aiming to limit memory usage to 1KB per image. Departing from the prevalent focus on performance enhancements, this work prioritizes the crucial trade-off between performance and memory requirements. The proposed model uses a transformer-based architecture designed to estimate image-to-image similarity by capturing interactions within and across images based on their local descriptors. A distinctive property of the model is the capability for asymmetric similarity estimation. Database images are represented with a smaller number of descriptors compared to query images, enabling performance improvements without increasing memory consumption. To ensure adaptability across different applications, a universal model is introduced that adjusts to a varying number of local descriptors during the testing phase. Results on standard benchmarks demonstrate the superiority of our approach over both hand-crafted and learned models. In particular, compared with current state-of-the-art methods that overlook their memory footprint, our approach not only attains superior performance but does so with a significantly reduced memory footprint. The code and pretrained models are publicly available at: https://github.com/pavelsuma/ames

In this work, we introduce OMG-Fuser, a fusion transformer-based network designed to extract information from various forensic signals to enable robust image forgery detection and localization. Our approach can operate with an arbitrary number of forensic signals and leverages object information for their analysis – unlike previous methods that rely on fusion schemes with few signals and often disregard image semantics. To this end, we design a forensic signal stream composed of a transformer guided by an object attention mechanism, associating patches that depict the same objects. In that way, we incorporate object-level information from the image. Each forensic signal is processed by a different stream that adapts to its peculiarities. A token fusion transformer efficiently aggregates the outputs of an arbitrary number of network streams and generates a fused representation for each image patch. We assess two fusion variants on top of the proposed approach: (i) score-level fusion that fuses the outputs of multiple image forensics algorithms and (ii) feature-level fusion that fuses low-level forensic traces directly. Both variants exceed state-of-the-art performance on seven datasets for image forgery detection and localization, with a relative average improvement of 12.1% and 20.4% in terms of F1. Our model is robust against traditional and novel forgery attacks and can be expanded with new signals without training from scratch. Our code is publicly available at: https://github.com/mever-team/omgfuser

The video object segmentation (VOS) task involves the segmentation of an object over time based on a single initial mask. Current state-of-the-art approaches use a memory of previously processed frames and rely on matching to estimate segmentation masks of subsequent frames. Lacking any adaptation mechanism, such methods are prone to test-time distribution shifts. This work focuses on matching-based VOS under distribution shifts such as video corruptions, stylization, and sim-to-real transfer. We explore test-time training strategies that are agnostic to the specific task as well as strategies that are designed specifically for VOS. This includes a variant based on mask cycle consistency tailored to matching-based VOS methods. The experimental results on common benchmarks demonstrate that the proposed test-time training yields significant improvements in performance. In particular for the sim-to-real scenario and despite using only a single test video, our approach manages to recover a substantial portion of the performance gain achieved through training on real videos. Additionally, we introduce DAVIS-C, an augmented version of the popular DAVIS test set, featuring extreme distribution shifts like image-/video-level corruptions and stylizations. Our results illustrate that test-time training enhances performance even in these challenging cases.

We introduce S^2VS, a video similarity learning approach with self-supervision. Self-Supervised Learning (SSL) is typically used to train deep models on a proxy task so as to have strong transferability on target tasks after fine-tuning. Here, in contrast to prior work, SSL is used to perform video similarity learning and address multiple retrieval and detection tasks at once with no use of labeled data. This is achieved by learning via instance-discrimination with task-tailored augmentations and the widely used InfoNCE loss together with an additional loss operating jointly on self-similarity and hard-negative similarity. We benchmark our method on tasks where video relevance is defined with varying granularity, ranging from video copies to videos depicting the same incident or event. We learn a single universal model that achieves state-of-the-art performance on all tasks, surpassing previously proposed methods that use labeled data.

In this paper, we address the problem of high performance and computationally efficient content-based video retrieval in large-scale datasets. Current methods typically propose either: (i) fine-grained approaches employing spatio-temporal representations and similarity calculations, achieving high performance at a high computational cost or (ii) coarse-grained approaches representing/indexing videos as global vectors, where the spatio-temporal structure is lost, providing low performance but also having low computational cost. In this work, we propose a Knowledge Distillation framework, called Distill-and-Select (DnS), that starting from a well-performing fine-grained Teacher Network learns: a) Student Networks at different retrieval performance and computational efficiency trade-offs and b) a Selector Network that at test time rapidly directs samples to the appropriate student to maintain both high retrieval performance and high computational efficiency. We train several students with different architectures and arrive at different trade-offs of performance and efficiency, i.e., speed and storage requirements, including fine-grained students that store/index videos using binary representations. Importantly, the proposed scheme allows Knowledge Distillation in large, unlabelled datasets – this leads to good students. We evaluate DnS on five public datasets on three different video retrieval tasks and demonstrate a) that our students achieve state-of-the-art performance in several cases and b) that the DnS framework provides an excellent trade-off between retrieval performance, computational speed, and storage space. In specific configurations, the proposed method achieves similar mAP with the teacher but is 20 times faster and requires 240 times less storage space. The collected dataset and implementation are publicly available at https://github.com/mever-team/distill-and-select.

In this paper we introduce ViSiL, a Video Similarity Learning architecture that considers fine-grained Spatio-Temporal relations between pairs of videos – such relations are typically lost in previous video retrieval approaches that embed the whole frame or even the whole video into a vector descriptor before the similarity estimation. By contrast, our Convolutional Neural Network (CNN)-based approach is trained to calculate video-to-video similarity from refined frame-to-frame similarity matrices, so as to consider both intra- and inter-frame relations. In the proposed method, pairwise frame similarity is estimated by applying Tensor Dot (TD) followed by Chamfer Similarity (CS) on regional CNN frame features - this avoids feature aggregation before the similarity calculation between frames. Subsequently, the similarity matrix between all video frames is fed to a four-layer CNN, and then summarized using Chamfer Similarity (CS) into a video-to-video similarity score – this avoids feature aggregation before the similarity calculation between videos and captures the temporal similarity patterns between matching frame sequences. We train the proposed network using a triplet loss scheme and evaluate it on five public benchmark datasets on four different video retrieval problems where we demonstrate large improvements in comparison to the state of the art. The implementation of ViSiL is publicly available.

This paper introduces the problem of Fine-grained Incident Video Retrieval (FIVR). Given a query video, the objective is to retrieve all associated videos, considering several types of associations that range from duplicate videos to videos from the same incident. FIVR offers a single framework that contains several retrieval tasks as special cases. To address the benchmarking needs of all such tasks, we construct and present a large-scale annotated video dataset, which we call FIVR-200K, and it comprises 225,960 videos. To create the dataset, we devise a process for the collection of YouTube videos based on major news events from recent years crawled from Wikipedia and deploy a retrieval pipeline for the automatic selection of query videos based on their estimated suitability as benchmarks. We also devise a protocol for the annotation of the dataset with respect to the four types of video associations defined by FIVR. Finally, we report the results of an experimental study on the dataset comparing five state-of-the-art methods developed based on a variety of visual descriptors, highlighting the challenges of the current problem.

This work addresses the problem of Near-Duplicate Video Retrieval (NDVR). We propose an efficient video-level NDVR scheme based on deep metric learning that leverages CNN features from intermediate layers to generate discriminative global video representations in tandem with a Deep Metric Learning (DML) framework with two fusion variations, trained to approximate an embedding function for accurate distance calculation between two near-duplicate videos. In contrast to most state-of-the-art methods, which exploit information deriving from the same source of data for both development and evaluation (which usually results to dataset-specific solutions), the proposed model is fed during training with sampled triplets generated from an independent dataset and is thoroughly tested on the widely used CC_WEB_VIDEO dataset. We demonstrate that the proposed approach achieves outstanding performance against the state-of-the-art, either with or without access to the evaluation dataset.

The large-scale availability of user-generated content in social media platforms has recently opened up new possibilities for studying and understanding the geospatial aspects of real-world phenomena and events. Yet, the large majority of user-generated content lacks proper geographic information (in the form of latitude and longitude coordinates). As a result, the problem of multimedia geotagging, i.e., extracting location information from user-generated text items when this is not explicitly available, has attracted increasing research interest. Here, we present a highly accurate geotagging approach for estimating the locations alluded by text annotations based on refined language models that are learned from massive corpora of social media annotations. We further explore the impact of different feature selection and weighting techniques on the performance of the approach. In terms of evaluation, we employ a large benchmark collection from the MediaEval Placing Task over several years. We demonstrate the consistently superior geotagging accuracy and low median distance error of the proposed approach using various data sets and comparing it against a number of state-of-the-art systems.