Other

NAVI is a category-agnostic multi-view image collection dataset with high-quality 3D scans and precise 2D–3D alignments (per-image camera poses). It was created to enable systematic evaluation of image-based 3D reconstruction, multi-view geometric correspondence, and surface-level/keypoint correspondence tasks from casual (in-the-wild) image collections where traditional SfM often fails. NAVI provides object-centric image collections paired with near-perfect ground-truth camera parameters and 3D shapes, enabling extraction of accurate cross-view correspondences and evaluation following protocols such as Probe3D. Primary resources: NAVI project site (https://navidataset.github.io/), NeurIPS 2023 paper and supplemental materials, and the google/navi GitHub repository.

No results tracked yet

SPair-71k is a large-scale benchmark dataset for semantic correspondence (semantic keypoint matching) introduced by Min et al. (2019). It contains 70,958 semantically paired images with large intra-class variations in viewpoint and scale and provides accurate, rich annotations intended for evaluating semantic correspondence methods. Annotations include per-image-pair semantic keypoint correspondences, bounding boxes, segmentation masks and metadata about viewpoint/scale variation, truncation and occlusion. The dataset is commonly used as a testbed for semantic keypoint/correspondence and matching algorithms and is distributed with a project page and an arXiv preprint (arXiv:1908.10543). A Hugging Face dataset mirror is also available.

No results tracked yet

Revisited Oxford (R-Oxford, also referred to as Roxford5k) is the corrected/re-annotated version of the classic Oxford Buildings image retrieval benchmark introduced in “Revisiting Oxford and Paris: Large-Scale Image Retrieval Benchmarking” (Radenović et al., CVPR 2018 / arXiv:1803.11285). The authors provide revised ground-truth annotations (including bounding boxes and an updated query list: the 55 original queries plus 15 new challenging queries = 70 queries), three evaluation protocols of different difficulty (Easy / Medium / Hard), and an optional R1M set of hard distractor images for large-scale testing. The “Medium” split is the medium-difficulty evaluation protocol from this benchmark (i.e., the dataset subset/protocol used when reporting Medium-difficulty mAP in papers). The dataset is widely used for instance-level image retrieval / landmark retrieval evaluation; the authors publish the images (original Oxford images) and the revisited annotation files (e.g. gnd_roxford5k.mat) and provide code and downloads from the project page.

No results tracked yet

Natural Questions (NQ) is a large question-answering corpus released by Google Research. Questions are real anonymized, aggregated queries issued to the Google search engine. For each question, annotators are given a selected Wikipedia page (from the top search results) and label a long answer (typically a paragraph) and, if present, a short answer (one or more entities or spans); pages can also be labeled null when no answer is present. NQ is intended to require reading and comprehension of entire Wikipedia articles and is used for open-domain / reading-comprehension QA research. The public release contains on the order of a few hundred thousand examples (commonly cited: ~307k training examples) and is English-only. License: CC BY-SA 3.0.

No results tracked yet