NYC-Indoor-VPR
NYC‑Indoor‑VPR is a unique indoor visual place recognition dataset created by New York University, comprising over 36,000 images captured from 13 crowded scenes across New York City under varying lighting and appearance conditions. The dataset was built using a semi‑automatic labeling method to establish ground truth locations for each image. It is primarily intended for indoor localization and navigation research, especially for robots and assistive navigation systems, addressing the challenges of visual aliasing and occlusion in indoor environments.
Dataset description and usage context
Dataset Overview
The NYC‑Indoor‑VPR dataset consists of synthetic and real data.
-
Synthetic Dataset
- Contains fisheye color images, depth maps, and 3D joint data.
- Approximately 233,000 image‑joint pairs collected from a Unity virtual environment.
- Download link
-
Real Dataset
- Contains fisheye color images and 3D joint data captured with a FLIR CM3‑U3‑13Y3C‑CS machine‑vision camera; joint data recorded by a Leap Motion device.
- 42,000 images collected from four participants.
- Intended for network fine‑tuning; primary training should use the synthetic dataset.
- Access request link
Data Format
- Joint Data
- Each frame's joint data is stored in a single .txt file.
- Files contain 3 lines of 21 numbers, comma‑separated.
- Every three consecutive numbers represent the x, y, z coordinates of a joint.
- The origin is the fisheye camera.
Usage Terms
- The dataset is provided free of charge for academic research by educational or research institutions for non‑commercial purposes.
- Downloading the dataset implies agreement not to redistribute, modify, or use it commercially.
- Publications using the dataset must cite the relevant papers.
Pair the dataset with AI analysis and content workflows.
Once the source passes your review, move straight into summarization, transformation, report drafting, or presentation generation with the JuheAI toolchain.