Grand Challenge
Objective Quality Assessment Methods for Light field Coding Applications
Abstract: Light field imaging has emerged as a transformative technology for capturing and visualizing three-dimensional scenes. By recording spatial and angular information, light fields enable advanced applications in immersive media, virtual reality, computational photography, and scientific imaging. However, due to the massive amount of data generated, efficient compression techniques are essential for enabling practical use and widespread adoption of light field technology. Compression algorithms often introduce distortions, which can degrade visual quality and thus, assessing the visual impact of distortions is critical to ensuring that compressed light fields maintain their perceptual fidelity. The subjective quality assessment relies on human observers to assess perceptual quality. While these methods are highly reliable, they are time-consuming and impractical for large-scale or iterative testing scenarios. To overcome these limitations, objective quality assessment metrics have gained importance as they offer an automated, scalable, and repeatable approach to evaluating visual quality. This Grand Challenge seeks proposals for full-reference and no-reference objective light field quality assessment methodologies to evaluate perceptual quality in light field coding applications. These methodologies will be tested on subjectively annotated light field datasets to ensure alignment with human opinions.
Website: TBA
Urban Elements ReID Challenge
Abstract: The objective of this competition is to encourage the development of algorithms for long-term video-based re-identification of urban elements, specifically focusing on trash bins, waste containers, and crosswalks. The proposed database is a continuation of the work carried out for the UrbAM-ReID dataset, published in ICIP 2024. The new dataset has been recorded in a different setting, specifically in an urban environment instead of a university campus. The recordings were captured using an on-board camera mounted on a vehicle, which allowed for the acquisition of urban elements during the route. It consists of images of crosswalks, containers and rubbish bins extracted from four different videos of the same route through an unspecified city in different directions (three recordings in one direction and one in the opposite, used for evaluation). To ensure that this database is as realistic as possible and provides sufficient variability for this type of task, the videos were recorded over four months, thus covering different lighting conditions and seasons of the year. This challenge aims to promote the use of intelligent systems for the automated monitoring of urban infrastructure, enabling more efficient and sustainable management through Computer Vision and Artificial Intelligence.
Beyond Visible Spectrum: AI for Sustainable Agriculture
Abstract: The “Beyond Visible Spectrum: AI for Sustainable Agriculture” challenge merges hyperspectral imaging with AI to enhance precision farming. It invites global researchers, experts, and students to address crop disease detection and agricultural data scarcity. Two main tasks defined in the challenge. The first requires participants to develop deep learning models for classifying diseased and healthy crops using real hyperspectral imagery (450-950 nm) captured by UAVs. These leverages detailed spectral data for early disease detection, aiming to reduce yield losses. The second task focuses on generating synthetic hyperspectral data using models like GANs, VAEs, and diffusion models. This synthetic data will augment limited real datasets, improving AI model robustness and generalization by simulating diverse crop conditions. Building on previous success (190 participants, 126 submissions last year), the challenge provides pre-processed, high-resolution datasets. It acts as a collaborative platform bridging AI research and practical agriculture, driving advancements in crop monitoring and synthetic data generation. The initiative ultimately seeks to contribute to global food security and sustainable farming practices, fostering a more resilient agricultural future.
Digit Recognition – Low Power and Speed Challenge
Abstract: Compare your low-power/high-speed FPGA digit recognition design with the best in the world! Hardware acceleration for machine-learning-based image recognition has become a very important topic in recent years. Especially when using FPGAs, one can obtain fast detection rates at high reliabilities without large infrastructure investments (as compared to an ASIC design) and still maintain low energy requirements. It allows for bringing fast and reliable image detection to edge devices. This has several advantages ranging from enhanced privacy, no communication overhead, lower latency, and lower energy requirements compared to cloud-based processing. The aim of the challenge is to compare recent developments in FPGA-based hardware accelerators in terms of inference speed and low power while still providing high accuracy rates. For this, the MNIST dataset of handwritten digits is used. MNIST is a widely recognized dataset that allows for easy comparison of different machine learning approaches. You are invited to submit your FPGA design results. The fastest design requiring the least energy (see link for details) wins!
Poison Sample Detection and Trigger Retrieval in Multimodal VQA Models
Abstract: The proliferation of Vision-Language Models (VLMs) across critical sectors such as healthcare, defense, and autonomous systems has brought remarkable advancements—but also heightened concerns about their vulnerability to backdoor attacks. These attacks introduce hidden triggers during training that, when activated, lead to adversarial model behavior. This grand challenge seeks to address the pressing issue of backdoor detection and mitigation in VLMs through a structured evaluation of poisoned input samples. Participants will analyze poisoned versus clean samples and assess the severity of injected triggers using a provided image dataset. The challenge promotes a deeper understanding of how adversaries can exploit seemingly benign data to compromise high-stakes AI systems. By encouraging innovative detection strategies and collaborative experimentation, the challenge aims to advance the development of robust, secure, and trustworthy VLMs. Outcomes from this competition will inform best practices and future standards in AI security, ensuring the safe deployment of VLMs in real-world environments where reliability is non-negotiable. Through this initiative, we aim to foster resilience in next-generation AI models and safeguard against emerging threats in our increasingly interconnected and AI-dependent world
Website: https://jj-vice.github.io/
Cityscape Aerial Image Dataset for Object Detection
Abstract: This challenge on CADOT (Cityscape Aerial image Dataset for Object deTection) calls on participants to design or apply cutting-edge learning-based object detection techniques tailored to optical remote sensing images. At its core is a novel, meticulously annotated dataset centered on one of the departments of Paris region, featuring 14 distinct object categories. The high-resolution raw images from which we built CADOT are those provided by the National Institute of Geographic and Forest Information (IGN), a public administrative establishment in France under the joint authority of the ministries responsible for ecology and forests. This dataset is specifically designed to address the complexities of dense urban environments, where high object density and unique urban characteristics present significant detection challenges. Participants will develop and evaluate advanced object detection algorithms, and are encouraged to explore generative AI for data augmentation. This is especially critical given the inherent imbalance of the dataset, which includes a high proportion of small objects, a common issue in remote sensing datasets. We provide comprehensive training and validation datasets, along with unannotated test dataset. We encourage innovative solutions that push the boundaries of object detection in urban remote sensing, fostering breakthroughs in this rapidly evolving field.
Website: https://cadot.onrender.com
Colorectal Cancer Tumor Grade Segmentation in Digital Histopathology Images: From Giga to Mini Challenge
Abstract: Colorectal cancer (CRC) is the third most prevalent cancer globally, affecting over 1.8 million new cases annually. It is also the second leading cause of cancer-related deaths worldwide in terms of death rates. By 2043, the number of CRC cases is projected to reach 3.2 million globally. Due to its complex pathophysiology, CRC has several subtypes that influence prognosis and treatment response. Colon cancer biopsies, obtained through colonoscopy or surgical excision, are routinely analyzed as part of histopathological evaluations. Distinguishing between benign and malignant tumors, as well as determining the tumor grade, are critical tasks for pathologists in their daily practice. Identifying the tumor grade is crucial, as it correlates strongly with patient prognosis, with poor differentiation linked to worse outcomes, and plays a key role in determining appropriate treatment options. In this challenge, a dataset of 103 digital histopathology whole slide images (WSI) collected from 103 patients with varying magnifications levels will be used. WSIs were pixelwise annotated by expert pathologists into 5 classes: tumor grades 1 through 3, normal mucosa and others. Submissions (through CodaLab) will be judged based on their macro F-score obtained over five classes. The dataset contains large SVS files, and their downsized versions as well. We offer downscaled version and original SVS files as training data to the competitors.