Few-Shot Learning

The Paris-Lille-3D is a Benchmark on Point Cloud Classification. The Point Cloud has been labeled entirely by hand with 50 different classes. The dataset consists of around 2km of Mobile Laser System point cloud acquired in two cities in France (Paris and Lille). Source: Paris-Lille-3D: a large and high-quality ground truth urban point cloud dataset for automatic segmentation and classification Image Source: https://npm3d.fr/paris-lille-3d

MedConceptsQA - Open Source Medical Concepts QA Benchmark The benchmark can be found here: https://huggingface.co/datasets/ofir408/MedConceptsQA

FSDKaggle2018 is an audio dataset containing 11,073 audio files annotated with 41 labels of the AudioSet Ontology. FSDKaggle2018 has been used for the DCASE Challenge 2018 Task 2. All audio samples are gathered from Freesound and are provided as uncompressed PCM 16 bit, 44.1 kHz mono audio files. The 41 categories of the AudioSet Ontology are: "Acoustic_guitar", "Applause", "Bark", "Bass_drum", "Burping_or_eructation", "Bus", "Cello", "Chime", "Clarinet", "Computer_keyboard", "Cough", "Cowbell", "Double_bass", "Drawer_open_or_close", "Electric_piano", "Fart", "Finger_snapping", "Fireworks", "Flute", "Glockenspiel", "Gong", "Gunshot_or_gunfire", "Harmonica", "Hi-hat", "Keys_jangling", "Knock", "Laughter", "Meow", "Microwave_oven", "Oboe", "Saxophone", "Scissors", "Shatter", "Snare_drum", "Squeak", "Tambourine", "Tearing", "Telephone", "Trumpet", "Violin_or_fiddle", "Writing". Source: https://zenodo.org/record/2552860 Image Source: https://labs.freesound.org/datasets/

TACO is a growing image dataset of waste in the wild. It contains images of litter taken under diverse environments: woods, roads and beaches. These images are manually labelled and segmented according to a hierarchical taxonomy to train and evaluate object detection algorithms. The annotations are provided in COCO format. Source: https://github.com/pedropro/TACO Image Source: https://github.com/pedropro/TACO

The MedNLI dataset consists of the sentence pairs developed by Physicians from the Past Medical History section of MIMIC-III clinical notes annotated for Definitely True, Maybe True and Definitely False. The dataset contains 11,232 training, 1,395 development and 1,422 test instances. This provides a natural language inference task (NLI) grounded in the medical history of patients. Source: MT-Clinical BERT: Scaling Clinical Information Extraction with Multitask Learning Image Source: https://arxiv.org/abs/1904.02181

GINC (Generative In-Context learning Dataset) is a small-scale synthetic dataset for studying in-context learning. The pretraining data is generated by a mixture of HMMs and the in-context learning prompt examples are also generated from HMMs (either from the mixture or not). The prompt examples are out-of-distribution with respect to the pretraining data since every example is independent, concatenated, and separated by delimiters. The GitHub repository provides code to generate GINC-style datasets of varying vocabulary sizes, number of HMMs, and other parameters.

ORBIT is a real-world few-shot dataset and benchmark grounded in a real-world application of teachable object recognizers for people who are blind/low vision. The dataset contains 3,822 videos of 486 objects recorded by people who are blind/low-vision on their mobile phones, and the benchmark reflects a realistic, highly challenging recognition problem, providing a rich playground to drive research in robustness to few-shot, high-variation conditions.

Collected from top 10 most popular clothing/wearable brandname logos captured in rich visual context. Source: Deep Learning Logo Detection with Data Expansion by Synthesising Context

N-Digit MNIST is a multi-digit MNIST-like dataset.

Baseline code for the three tracks of ExVo 2022 competition. Consists of 59,201 recordings totaling more than 36 hours of audio data from 1 702 speakers. To our knowledge, this is substantially larger than than any previously available dataset of human vocal bursts.

F-SIOL-310 is a robotic dataset and benchmark for Few-Shot Incremental Object Learning, which is used to test incremental learning capabilities for robotic vision from a few examples. A robot was used to actively capture household objects on a table. The dataset is specifically designed for FSIL with only a small set of training images and a larger set of test images per object category captured by the robot using its own camera and it considers various other robot vision challenges as well, such as different object sizes, object transparency and a clear distinction between objects in the train and test sets. It contains images of 310 objects from 22 categories.

The Few-Shot Object Learning (FewSOL) dataset can be used for object recognition with a few images per object. It contains 336 real-world objects with 9 RGB-D images per object from different views. Object segmentation masks, object poses and object attributes are provided. In addition, synthetic images generated using 330 3D object models are used to augment the dataset. FewSOL dataset can be used to study a set of few-shot object recognition problems such as classification, detection and segmentation, shape reconstruction, pose estimation, keypoint correspondences and attribute recognition. Motivation: If robots can recognize objects from a few exemplar images, it is possible to scale up the number of objects a robot can recognize because collecting a few images per object is a much easier process compared to building a 3D model of an object. In addition, models trained in the meta-learning setting can generalize to new objects without re-training.

The FIGR-8 database is a dataset containing 17,375 classes of 1,548,256 images representing pictograms, ideograms, icons, emoticons or object or conception depictions. Its aim is to set a benchmark for Few-shot Image Generation tasks, albeit not being limited to it. Each image is represented by 192x192 pixels with grayscale value of 0-255. Classes are not balanced (they do not all contain the same number of elements), but they all do contain at the very least 8 images. Source: https://github.com/marcdemers/FIGR-8 Image Source: https://github.com/marcdemers/FIGR-8

N-Omniglot is a neuromorphic dataset for few-shot learning. It contains 1,623 categories of handwritten characters, with only 20 samples per class.

Contains 3,689,229 English news articles on politics, gathered from 11 United States (US) media outlets covering a broad ideological spectrum.

Due to the highly variable sample size of the original BirdClef2020 dataset and the issues that it presents with reproducibility, we propose a pruned version of the set, where samples longer than 180s are removed along with classes with fewer than 50 samples. This processing puts it further in line with other complex audio datasets and allows for experimentation on more consumer friendly hardware.

Bongard-OpenWorld is a new benchmark for evaluating real-world few-shot reasoning for machine vision. We hope it can help us better understand the limitations of current visual intelligence and facilitate future research on visual agents with stronger few-shot visual reasoning capabilities.

Intended to provide freely available data sets in various formats together with basic annotation to be useful for applications in computational linguistics, translation studies and cross-linguistic corpus studies. Source: Parallel Data, Tools and Interfaces in OPUS

FREDo is a Few-Shot Document-Level Relation Extraction Benchmark based on DocRED and SciERC. The dataset is divided into four subsets: training set (62 relations), validation set (16 relations), in-domain test set (16 relations), and cross-domain test set (7 relations).

GAS (Grasp Area Segmentation) dataset consists of 10089 RGB images of cluttered scenes grouped into 1121 grasp-area segmentation tasks. For each RGB image we provide a binary segmentation map with the graspable and non-graspable regions for every object in the scene. The dataset can be used for meta-training part-based grasp area estimation networks. For creating the GAS dataset we use the RGB images and corresponding ground truth segmentation masks from the GraspNet 1-Billion dataset.

carecall is a Korean dialogue dataset for role-satisfying dialogue systems. The dataset was composed with a few samples of human-written dialogues using in-context few-shot learning of large-scale LMs. Large-scale LMs can generate dialogues with a specific personality, given a prompt consisting of a brief description of the chatbot’s properties and few dialogue examples. We use this method to build the entire dataset.

"We built a large lung CT scan dataset for COVID-19 by curating data from 7 public datasets listed in the acknowledgements. These datasets have been publicly used in COVID-19 diagnosis literature and proven their efficiency in deep learning applications. Therefore, the merged dataset is expected to improve the generalization ability of deep learning methods by learning from all these resources together. These datasets are made available in different formats. Our goal is to provide a large dataset of COVID-19, Normal, and CAP CT slices together with their corresponding metadata. Some of the datasets consist of categorized CT slices, and some include CT volumes with annotated lesion slices. Therefore, we used the slice-level annotations to extract axial slices from CT volumes. We then converted all the images to 8-bit to have a consistent depth. To ensure the dataset quality, we have removed the closed lung normal slices that do not carry information about inside lung manifestations. Additionally, we did not include images lacking clear class labels or patient information. In total, we have gathered 7,593 COVID-19 images from 466 patients, 6,893 normal images from 604 patients, and 2,618 CAP images from 60 patients. All of our CAP images are from Afshar et al. dataset, in which 25 cases are already annotated. Our radiologist has annotated the remaining 35 CT scan volumes. This is the largest COVID-19 lung CT dataset so far, to the best of our knowledge." - Source: A Robust Ensemble-Deep Learning Model for COVID-19 Diagnosis based on an Integrated CT Scan Images Database Acknowledgements - J. Zhao, Y. Zhang, X. He, and P. Xie, "COVID-CT-Dataset: a CT scan dataset about COVID-19," arXiv preprint arXiv:2003.13865, 2020. - P. Afshar et al., "COVID-CT-MD: COVID-19 Computed Tomography (CT) Scan Dataset Applicable in Machine Learning and Deep Learning," arXiv preprint arXiv:2009.14623, 2020. - J. P. Cohen, P. Morrison, L. Dao, K. Roth, T. Q. Duong, and M. Ghassemi, "Covid-19 image data collection: Prospective predictions are the future," arXiv preprint arXiv:2006.11988, 2020. - S. Morozov et al., "MosMedData: Chest CT Scans With COVID-19 Related Findings Dataset," arXiv preprint arXiv:2005.06465, 2020. - M. Rahimzadeh, A. Attar, and S. M. Sakhaei, "A Fully Automated Deep Learning-based Network For Detecting COVID-19 from a New And Large Lung CT Scan Dataset," medRxiv, 2020. - M. Jun et al., "COVID-19 CT Lung and Infection Segmentation Dataset," Zenodo, Apr, vol. 20, 2020. - "COVID-19." 2020. [Online] http://medicalsegmentation.com/covid19/ [Accessed 23 December, 2020].

The speech accent archive uniformly presents a large set of speech samples from a variety of language backgrounds. Native and non-native speakers of English read the same paragraph and are carefully transcribed. The archive is used by people who wish to compare and analyze the accents of different English speakers.

ToM-in-AMC is a novel NLP benchmark, Short for Theory-of-Mind meta-learning Assessment with Movie Characters. The benchmark consists of 1,000 parsed movie scripts for this purpose, each corresponding to a few-shot character understanding task. Source: Few-Shot Character Understanding in Movies as an Assessment to Meta-Learning of Theory-of-Mind Image Source: https://arxiv.org/pdf/2211.04684v1.pdf

One of the founding fathers of marine mammal bioacoustics, William Watkins, carried out pioneering work with William Schevill at the Woods Hole Oceanographic Institution for more than four decades, laying the groundwork for our field today. One of the lasting achievements of his career was the Watkins Marine Mammal Sound Database, a resource that contains approximately 2000 unique recordings of more than 60 species of marine mammals (Table 1). Recordings were made by Watkins and Schevill as well as many others, including G. C. Ray, D. Wartzok, D. and M. Caldwell, K. Norris, and T. Poulter. Most of these have been digitized, along with approximately 15,000 annotated digital sound clips. The Watkins database has enormous historical and scientific value. The recordings provide sounds professionally identified as produced by particular marine mammal species in defined geographic regions (Figure 1) during specific seasons, which can be used as reference datasets for marine mammal detections from the growing amounts of passive acoustic monitoring (PAM) data that are being collected worldwide. In addition, the archive contains recordings that span seven decades, from the 1940's to the 2000's, and includes the very first recordings of 51 species of marine mammals. These data provide a rich resource to efforts aimed at examining long-term changes in vocal production that may be related to changes in ambient noise levels, as well as serve as a voucher collection for many species. We have made this resource fully accessible online, as was Watkins' goal (see crediting information below). The final product enables investigators, educators, students, and the public worldwide to freely and easily access acoustic samples from identified species of marine mammals, and place these samples in a geographic and temporal context. The physical collection has been donated to the New Bedford Whaling Museum.

Millions of people around the world have low or no vision. Assistive software applications have been developed for a variety of day-to-day tasks, including currency recognition. To aid with this task, we present BankNote-Net, an open dataset for assistive currency recognition. The dataset consists of a total of 24,816 embeddings of banknote images captured in a variety of assistive scenarios, spanning 17 currencies and 112 denominations. These compliant embeddings were learned using supervised contrastive learning and a MobileNetV2 architecture, and they can be used to train and test specialized downstream models for any currency, including those not covered by our dataset or for which only a few real images per denomination are available (few-shot learning). We deploy a variation of this model for public use in the last version of the Seeing AI app developed by Microsoft, which has over a 100 thousand monthly active users.

A Zero-Shot Sketch-based Inter-Modal Object Retrieval Scheme for Remote Sensing Images WITH the advancement in sensor technology, huge amounts of data are being collected from various satellites. Hence, the task of target-based data retrieval and acquisition has become exceedingly challenging. Existing satellites essentially scan a vast overlapping region of the Earth using various sensing techniques, like multi-spectral, hyperspectral, Synthetic Aperture Radar (SAR), video, and compressed sensing, to name a few. With increasing complexity and different sensing techniques at our disposal, it has become our primary interest to design efficient algorithms to retrieve data from multiple data modalities, given the complementary information that is captured by different sensors. This type of problem is referred to as inter-modal data retrieval. In remote sensing (RS), there are primarily two important types of problems, i.e., land-cover classification and object detection. In this work, we focus on the target-based object retrieval part, which falls under the realm of object detection in RS. Object retrieval essentially requires high-resolution imagery for objects to be distinctly visible in the image. The main challenge with the conventional retrieval approach using large-scale databases is that, quite often, we do not have any query image sample of the target class at our disposal. The target of interest solely exists as a perception to the user in the form of an imprecise sketch. In such situations where a photo query is absent, it can be immensely useful if we can promptly make a quick hand-made sketch of the target. Sketches are a highly symbolic and hieroglyphic representation of data. One can exploit the notion of this minimalistic representative of sketch queries for sketch-based image retrieval (SBIR) framework. While dealing with satellite images, it is imperative to collect as many samples of images as possible for each object class for object recognition with a high success rate. However, in general, there exists a considerable number of classes for which we seldom have any training data samples. Therefore, for such classes, we can use the zero-shot learning (ZSL) strategy. The ZSL approach aims to solve a task without receiving any example of that task during the training phase. This makes the network capable of handling an unseen class (new class) sample obtained during the inference phase upon deployment of the network. Hence, we propose the aerial sketch-image dataset, namely Earth on Canvas dataset. Classes in this dataset: Airplane, Baseball Diamond, Buildings, Freeway, Golf Course, Harbor, Intersection, Mobile home park, Overpass, Parking lot, River, Runway, Storage tank, Tennis court.

Introduction The FewGLUE_64_labeled dataset is a new version of FewGLUE dataset. It contains a 64-sample training set, a development set (the original SuperGLUE development set), a test set, and an unlabeled set. It is constructed to facilitate the research of few-shot learning for natural language understanding tasks. Compared with the original FewGLUE dataset, it differs in the number of labeled data examples in the training set, where the original FewGLUE has 32 training examples while FewGLUE_64_labeled has 64 labeled examples. Purposes for constructing a new version of FewGLUE dataset include: 1. To answer the questions that what is the best performance that few-shot learning can achieve and whether it is possible to further close the performance gap between few-shot learning and fully-supervised systems. 2. To explore to which degree the number of labeled training examples influences the few-shot performance. Please refer to the FewNLU paper as well as the FewNLU leaderboard for more details. ### Acknowledgement Part of the FewGLUE_64_labeled dataset is based on the original 32-sample version of FewGLUE. We collect them together in one package for the convenience of usage. We appreciate all the contributors who made their dataset public, which greatly advanced few-shot learning as well as the FewNLU project.

ISEKAI dataset’s images are generated by Midjourney’s text-to-image model using well-crafted instructions. Images were manually selected to ensure core concept consistency. The dataset currently comprises 20 groups, and 40 categories in total (continues to grow). Each group pairs a new concept with a related real-world concept, like "octopus vacuum" and "octopus." These can serve as challenging negative samples for each other. Each concept has no less than 32 images, supporting multi-shot examples.

MatSim is a synthetic dataset, and natural image benchmark for computer vision-based recognition of similarities and transitions between materials and textures, focusing on identifying any material under any conditions using one or a few examples (one-shot learning), including materials states and subclasses. Source: One-shot recognition of any material anywhere using contrastive learning with physics-based rendering Image Source: https://arxiv.org/pdf/2212.00648v1.pdf

A large-scale reference dataset for bioacoustics. MeerKAT is a 1068h large-scale dataset containing data from audio-recording collars worn by free-ranging meerkats (Suricata suricatta) at the Kalahari Research Centre, South Africa, of which 184h are labeled with twelve time-resolved vocalization-type ground truth target classes, each with millisecond resolution. The labeled 184h MeerKAT subset exhibits realistic sparsity conditions for a bioacoustic dataset (96% background-noise or other signals and 4% vocalizations), dispersed across 66398 10-second samples, spanning 251562 labeled events and showcasing significant spectral and temporal variability, making it the first large-scale reference point with real-world conditions for benchmarking pretraining and finetune approaches in bioacoustics deep learning.

Dataset release for the BMVC 2021 Paper "Few-Shot Domain Adaptation for Low Light RAW Image Enhancement" Abstract: Enhancing practical low light raw images is a difficult task due to severe noise and color distortions from short exposure time and limited illumination. Despite the success of existing Convolutional Neural Network (CNN) based methods, their performance is not adaptable to different camera domains. In addition, such methods also require large datasets with short-exposure and corresponding long-exposure ground truth raw images for each camera domain, which is tedious to compile. To address this issue, we present a novel few-shot domain adaptation method to utilize the existing source camera labeled data with few labeled samples from the target camera to improve the target domain’s enhancement quality in extreme low-light imaging. Our experiments show that only ten or fewer labeled samples from the target camera domain are sufficient to achieve similar or better enhancement performance than training a model with a large labeled target camera dataset. To support research in this direction, we also present a new low-light raw image dataset captured with a Nikon camera, comprising short-exposure and their corresponding long-exposure ground truth images. The code is available at https://val.cds.iisc.ac.in/HDR/BMVC21/index.html.

Open MIC (Open Museum Identification Challenge) contains photos of exhibits captured in 10 distinct exhibition spaces of several museums which showcase paintings, timepieces, sculptures, glassware, relics, science exhibits, natural history pieces, ceramics, pottery, tools and indigenous crafts. The goal of Open MIC is to stimulate research in domain adaptation, egocentric recognition and few-shot learning by providing a testbed complementary to the famous Office 31. Source: Museum Exhibit Identification Challenge for Domain Adaptation and Beyond

Green family of datasets for emergent communications on relations. By comparison with other relations datasets, TexRel provides rapid training and experimentation, whilst being sufficiently large to avoid overfitting in the context of emergent communications.

A dataset specifically tailored to the biotech news sector, aiming to transcend the limitations of existing benchmarks. This dataset is rich in complex content, comprising various biotech news articles covering various events, thus providing a more nuanced view of information extraction challenges. A standout feature of our dataset is its dual-layered approach: it not only identifies the general theme of the articles but also extracts specific information about the target companies involved. This dual focus significantly elevates the dataset’s relevance for applications demanding a deeper insight into the interplay of events, companies, and the broader biotech industry. The dataset encompasses 31 classes, including a ‘None’ category, to cover various events and information types such as event organisation, executive statements, regulatory approvals, hiring announcements, and more. Key aspects: - Event extraction - Multi-class classification - Biotech news domain - 31 classes - 3140 total number of examples Classes: - alliance & partnership - forming an alliance or partnership - article publication - publishing an article - clinical trial sponsorship - Sponsoring or participating in a clinical trial - closing - shutting down a facility/office/division or ceasing an initiative - company description - describing or profiling the company - department establishment - establishing a new department or division - event organization - organizing or participating in an event like a conference, exhibition, etc - event organization - organizing or participating in an event - executive appointment - appointing a new executive - executive statement - a statement or quote from an executive of a company - expanding geography - expanding into new geographical areas - expanding industry - expanding into new industries or markets - foundation - establishing a new charitable foundation - funding round - raising a new round of funding - hiring - announcing new hires or appointments at the company - investment in public company - making an investment in a public company - ipo exit - having an initial public offering or acquisition exit - m&a - mergers, acquisitions, or divestitures - new initiatives & programs - announcing new initiatives or programs - new initiatives or programs - announcing new initiatives, programs, or campaigns - None - no label - other - other events that don't fit into defined categories - participation in an event - participating in an industry event, conference, etc - partnerships & alliances - forming partnerships or strategic alliances with other companies - patent publication - publication of a new patent filing - product launching & presentation - launching or unveiling a new product - product updates - announcing updates or new versions of existing products - regulatory approval - getting approval from regulatory bodies for products, services, trials, etc. - service & product providing - launching or expanding products or services - subsidiary establishment - establishing a new subsidiary company - support & philanthropy - philanthropic activities or donations