Datasets

Datasets and benchmarks released alongside our research, with links to the data, the papers that describe them, and citations.

If you use one of these datasets, please cite the paper that introduced it. BibTeX is provided with each dataset.

• Language & multimodalRecognition2026

  VisionCrowdTestCrowd Counting Benchmark for Vision-Language Models

  1,304 images398K+ annotated heads

  VisionCrowdTest is a diagnostic benchmark that asks whether the crowd counts produced by vision-language models are actually supported by spatial evidence in the image. It contains 1,304 images with more than 398K annotated heads, covering RGB, aerial, thermal and infrared imagery across low, medium and high crowd densities. Five recent VLMs are evaluated zero-shot on counting and spatial grounding, revealing a semantic-localization gap: plausible counts that correspond only weakly to where people actually are.

  PaperDo Vision-Language Models Truly Count? Revealing the Semantic-Localization Gap in Dense Crowd UnderstandingInternational Conference on Neural Information Processing (ICONIP), 2026

  GitHub

• Recognition2026

  GenSyn10Multi-Generator Synthetic Image Dataset

  60,000 images3 generators

  GenSyn10 is a synthetic counterpart to CIFAR-10: 60,000 images at 32×32 across the same ten classes, with the standard 50k/10k train/test split. The images were rendered from grammar-based prompts by three architecturally different text-to-image models (FLUX.2-dev, HunyuanImage-3.0 and Qwen-Image-2512), making it a controlled benchmark for detecting AI-generated images and for studying generalisation across generators. Seventeen classification models are evaluated under a four-stage protocol: real-data baseline, zero-shot transfer, fine-tuning and retention.

  PaperGenSyn10: A Multi-Generative AI Dataset For Benchmarking Image ClassificationarXiv preprint, 2026

  DatasetarXiv

• Language & multimodal2026

  DermCaseRare Skin Disease Diagnostic Reasoning Benchmark

  6,354 rare cases26,030 image–caption pairs

  DermCase is a long-context dermatology benchmark for testing whether large vision-language models can reason their way to a diagnosis of rare skin diseases. Built from publicly licensed, peer-reviewed case reports, it holds 6,354 rare cases with complete diagnostic reasoning chains and 26,030 dermatological image–caption pairs. It supports evaluation of diagnostic accuracy, differential diagnosis and multi-step clinical reasoning, and 22 LVLMs are benchmarked on it.

  PaperMind the Rarities: Can Rare Skin Diseases Be Reliably Diagnosed via Diagnostic Reasoning?arXiv preprint, 2026

  Project pagearXiv

• Recognition2026

  DRDDDiverse Road Damage Dataset

  1,500 images5 damage types

  DRDD targets road scenes in which several kinds of damage appear in the same frame, with about three damages per image on average. Its 1,500 images (1920×1440) were extracted from GoPro footage recorded in Oman under varying weather and times of day, checked by a civil engineer, and annotated with bounding boxes in Pascal VOC and COCO formats. Five classes are covered (alligator, block, longitudinal, pothole and transverse cracks), with occlusions, shadows and motion blur adding to the difficulty.

  PaperRDD4D: 4D Attention-Guided Road Damage Detection And ClassificationIEEE Access, 2026

  DatasetarXivCode

• Language & multimodal2025

  PENDULUMAssessing Sycophancy in MLLMs

  751 images6 domains

  Dataset contains 751 images, each paired with approximately 2,000 visual question–answer (VQA) examples. The images span six distinct domains, each presenting different visual and reasoning challenges. For each question, three prompt types are included: a baseline (equilibrium) prompt with no user influence, a positive-influence prompt that provides a correct hint aligned with the visual evidence, and a negative-influence prompt that is misleading or adversarial.

  PaperPENDULUM: A Benchmark for Assessing Sycophancy in Multimodal Large Language ModelsarXiv preprint, 2025

  DatasetInput PromptsarXiv

• Recognition2025

  IARDIntruder Activity Recognition Dataset

  69,000+ frames4 classes

  IARD contains four classes — Armed Intruder, Door Kick, Intruder Inside, and Lock Breaking. Together, these classes span more than 69,000 frames. The dataset was built from carefully selected YouTube videos, cropped and manually annotated into the target activity classes, with standardized 224×224 frames. IARD was built with strict ethical safeguards: all videos were sourced from public sources, faces were anonymized, no identifying metadata was retained, and any sensitive or inappropriate content was removed. The dataset follows responsible AI and privacy principles.

  PaperIARD: Intruder Activity Recognition Dataset for Threat DetectionACM International Conference on Information and Knowledge Management (CIKM), 2025

  DatasetOfficial PDF

• Recognition2025

  RHADReal-World Hazardous Activities Dataset

  Real-world videoIndoor hazards

  RHAD is a video dataset curated for recognising hazardous activities in real-world indoor settings, which existing activity datasets cover only thinly. It accompanies HazardNet, a hybrid model that combines local and global spatio-temporal representation modules, and the paper benchmarks a range of state-of-the-art activity recognition models on it to support safety monitoring and preventive interventions.

  PaperTowards Hazardous Activity Recognition for A Novel Real-World DatasetACM International Conference on Multimedia (ACM MM), 2025

  GitHubOfficial PDF

• Low-level vision2024

  HazeSpace2MHazy Image Dataset

  2M+ images10 haze levels

  A collection of over 2 million images designed to enhance dehazing through haze type classification. HazeSpace2M includes diverse scenes with 10 haze intensity levels, featuring Fog, Cloud, and Environmental Haze.

  PaperHazeSpace2M: A Dataset for Haze Aware Single Image DehazingACM International Conference on Multimedia (ACM MM), 2024

  DatasetOfficial PDF

• Low-level vision2024

  LoLI-StreetLow-light Images of Streets

  33k image pairs4K/8K real test set

  The training consists of 30k, while validation has 3k paired low and high-light images. Moreover, we collected high-resolution videos (4K/8K at 60fps) from various cities under low-light conditions, extracting and manually reviewing frames to create the Real Low-light Testset (RLLT) of our LoLI-Street dataset. We used Photoshop v25.0 to generate the synthetic images of our dataset.

  PaperLoLI-Street: Benchmarking Low-Light Image Enhancement and BeyondAsian Conference on Computer Vision (ACCV), 2024

  Dataset (Kaggle)Dataset (Google Drive)arXiv

• Language & multimodal2024

  DefAnDefinitive Answer Dataset

  75,000+ samples8 knowledge domains

  DefAn is a comprehensive evaluation benchmark dataset, with more than 75000 samples, designed to assess the hallucination tendencies of large language models (LLMs). The dataset is categorized into eight knowledge domains: Sports, Census Australia, Nobel Prize, Entertainment, World Organizations, QS Ranking, Conference Venue, and Math. The dataset is structured into two parts: public and hidden.

  PaperDefAn: Definitive Answer Dataset for LLM Hallucination EvaluationInformation, 2025

  DatasetOfficial PDFarXiv

• Recognition2024

  UAV-AWIDUAVs in Adverse Weather and Image Distortions

  3 distortion types3 severity levels

  UAV-AWID is a set of test benchmarks for measuring how drone detectors cope with degraded images: rain at three intensities (drizzle, heavy, torrential), and motion blur and artificial noise at three levels each (low, medium, high). It is released together with the Complex Background Dataset (CBD) and an augmented version (ACBD) for training. Benchmarks of YOLOv5, YOLOv8, YOLO-NAS, Faster R-CNN and RetinaNet show mAP falling by up to 77 points under severe conditions, and training on the augmented data substantially improves robustness.

  PaperImpact of Adverse Weather and Image Distortions on Vision-Based UAV Detection: A Performance Evaluation of Deep Learning ModelsDrones, 2024

  DatasetOfficial PDF

• Recognition2024

  CBDComplex Background Dataset for UAV Detection

  12,492 imagesUAVs & birds

  CBD supports drone detection against cluttered scenes. It brings together 12,492 images drawn from several public UAV datasets, spanning urban, semi-urban, indoor, outdoor and mountain scenes in sunny and cloudy conditions, with birds included as distractors and annotations provided in YOLO and JSON formats. The same paper introduces the Sky Background Dataset (SBD, 3,888 sky-only images) and a Rainy Test Set (RTS), which adds synthetic drizzle, heavy and torrential rain to 2,528 test images to measure how detectors degrade in bad weather.

  PaperInvestigation of UAV Detection in Images With Complex Backgrounds and Rainy ArtifactsIEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 2024

  DatasetOfficial PDFarXivCode

• Language & multimodal2023

  SSIDSequential Storytelling Image Dataset

  17,365 images13,892 stories

  Visual Storytelling Task (VST) takes a set of images as input and aims to generate a coherent story relevant to the input images. We collected and manually annotated images from publicly available documentaries, lifestyles, and movies. In summary, the SSID dataset has 17,365 images of 3,473 unique sets of five images. Each set of images is associated with four ground truths, resulting in 13,892 unique ground truths (i.e., written stories). Each ground truth is composed of five connected sentences written as a story.

  PaperSequential Vision to Language as Story: A Storytelling Dataset and BenchmarkingIEEE Access, 2023

  DatasetOfficial PDF

• 3D & depth2023

  SwanA Point Cloud Dataset

  Outdoor LiDAR3 annotation tasks

  A large-scale outdoor point cloud dataset captured with LiDAR sensor and annotated manually for semantic segmentation, instance segmentation and object detection.

  PaperSAT3D: Slot Attention Transformer for 3D Point Cloud Semantic SegmentationIEEE Transactions on Intelligent Transportation Systems, 2023

  DatasetOfficial PDF

• 3D & depth2023

  Perth-WA LocalizationA Point Cloud Dataset

  City-scale LiDAR map6DoF ground truth

  LiDAR 3D point cloud map of Perth, Western Australia is provided with ground truth locations and orientations.

  PaperSlice Transformer and Self-supervised Learning for 6DoF Localization in 3D Point Cloud MapsIEEE International Conference on Robotics and Automation (ICRA), 2023

  DatasetarXiv

• Recognition2022

  FGBRFine Grained Butterflies and Reef Fish Dataset

  6,438 images100 species

  FGBR contains two species, butterfly and reef fish. The butterfly dataset has 50 species, consisting of 2613 images, most of which are ecological. The reef fish also has 50 species. Most of the 3825 images in the reef fish dataset are ecological.

  PaperTowards Low-Cost Classification for Novel Fine-Grained DatasetsElectronics, 2022

  DatasetOfficial PDF

• Recognition2022

  ANUBISAustralian National University Benchmarking Indoor Skeleton

  102 actions66,232 clips

  ANUBIS is a large-scale skeleton action recognition benchmark with 102 action classes performed by 80 subjects, totalling 66,232 clips recorded with Azure Kinect sensors as 32-joint 3D skeletons. It adds systematic back-view coverage, multi-person interactions, fine-grained and aggressive actions, and contemporary social behaviours, and it remains far from saturated: the best reported top-1 accuracy is around 60%, against more than 90% on the NTU benchmarks.

  PaperRepresentation-Centric Survey of Supervised Skeletal Action Recognition and the New BenchmarkPattern Recognition, 2026

  DatasetOfficial PDFarXivCode

• Recognition2022

  DVLPDDiverse Vehicle and License Plates Dataset

  10k imagesVehicles & plates

  We collect a novel dataset that has 10k images of Pakistani vehicles. These vehicles are trucks, buses, vans, carry vans, and cars. There are two sub-classes of trucks called type1 trucks having a single axle and type2 trucks with double axles. After dataset collection, the pre-processing steps and ground truth generation are performed to train and test the CNN models.

  PaperVehicle and License Plate Recognition with Novel Dataset for Toll CollectionPattern Analysis and Applications, 2025

  DatasetarXiv

• Low-level visionUnderwater2021

  HICRDHeron Island Coral Reef Dataset

  11,676 imagesReal coral reef

  Heron Island Coral Reef Dataset (HICRD) contains 6003 low-quality images, 3673 good-quality images, and 2000 restored images. We use low-quality images and restored images as the unpaired training set (trainA + trainB). In contrast, the paired training set contains good-quality (trainA_paired) images and corresponding restored images (trainB_paired). The test set contains 300 good-quality images (testA) as well as 300 paired restored images (testB) as ground truth. All images are in 1842 x 980 resolution. The copyright belongs to CSIRO (Commonwealth Scientific and Industrial Research Organisation).

  PaperUnderwater Image Restoration via Contrastive Learning and a Real-world DatasetRemote Sensing, 2022

  DatasetOfficial PDFarXiv

• Recognition2021

  RRCDRoman Republican Coin Dataset

  228 motif classesRoman coins

  Based on Crawford’s work, we collect the most diverse and extensive image dataset of the reverse sides. For most of the Roman Republic coin classes, the ob- verse side depicts more discriminative information than the observe side. Our dataset has 228 motif classes, including 100 classes that are the main classes for training and testing, which we call the main dataset RRCD-Main. The images of the additional 128 classes constitute the disjoint test set, RRCD-Disjoint, which we allocate to assess the generalization ability of our models.

  PaperDeep Ancient Roman Republican Coin Classification via Feature Fusion and AttentionPattern Recognition, 2021

  DatasetOfficial PDFarXiv

• 3D & depth2021

  SMDSynthetic Monocular Depth

  Synthetic depthGeneration code

  Synthetic Monocular Depth generation code.

  PaperDeblur and deep depth from single defocus imageMachine Vision and Applications, 2021

  Official PDFarXivCode

• Low-level vision2020

  NCDNatural-Color Dataset

  723 images20 categories

  The Natural-Color Dataset (NCD) is an image colorization dataset where images are true to their colors. For example, a carrot will have an orange color in most images. Bananas will be either greenish or yellowish. It contains 723 images from the internet distributed in 20 categories. Each image has an object and a white background.

  PaperImage Colorization: A Survey and DatasetInformation Fusion, 2025

  Dataset (GrayScale Images)Dataset (Color GroundTruth)Official PDFarXiv

• Low-level visionUnderwater2019

  UWSDUnderWater Synthetic Dataset

  10 water typesSynthetic underwater

  To synthesize underwater image degradation datasets, we use the attenuation coefficients described in Table 1 for the different water types of oceanic and coastal classes (i.e., I, IA, IB, II, and III for open ocean waters, and 1, 3, 5, 7, and 9 for coastal waters). Type-I is the clearest and Type-III is the most turbid open ocean water. Similarly, for coastal waters, Type-1 is the clearest and Type-9 is the most turbid. We apply Eqs (1) and (2) (please check the paper) to build ten types of underwater image datasets by using the RGB-D NYU-v2 indoor dataset which consists of 1449 images. To improve the quality of datasets, we crop the original size (480x640) of NYU-v2 to 460x620. This dataset is for non-commercial use only. The size of each dataset is 1.2GB.

  PaperUnderwater Scene Prior Inspired Deep Underwater Image and Video EnhancementPattern Recognition, 2019

  DatasetOfficial PDFarXiv