This page contains the description and the link to a subsampled version of the original CURE-OR dataset. The goal of this project is to analyze the robustness of off-the-shelf recognition applications under multifarious challenging conditions, investigate the relationship between the recognition performance and image quality, and estimate the performance based on hand-crafted features as well as data-driven features. To achieve this goal, we introduced a large-scale, controlled, and multi-platform object recognition dataset CURE-OR, which stands for Challenging Unreal and Real Environments for Object Recognition. For more information about CURE-OR, please refer to our papers.
Challenging Conditions
Objects
Backgrounds
5 Backgrounds: White, 2D Living room, 2D Kitchen, 3D Living room, 3D Office
Devices
5 Devices: iPhone 6s, HTC One X, LG Leon, Logitech C920 HD Pro Webcam, Nikon D80
Orientations
5 Object orientations: Front, Left, Back, Right, Top
Challenging Conditions Generation
Python Imaging Library (PIL) version 4.2.1 and scikit-image version 0.13.0 are utilized to generate challenging conditions. Salt and pepper noise is synthesized with scikit-image and all other challenging conditions are simulated with PIL. The minimum and maximum parameter values for each challenge type except dirty lens 2 and grayscale are provided below, and parameter values are linearly spaced between the two for different challenge levels.
- Resize: downsample with bicubic interpolation; the size of a new image is determined by multiplying pixel dimensions of an original image by factors linearly spaced between 1 (exclusive) and 0.5 (inclusive)
- Underexposure: brightness control with factors between 0.4 and 0.08
- Overexposure: brightness control with factors between 1.4 and 6
- Gaussian blur: radius of blur between 8 and 60
- Contrast: increase separation between dark and bright colors on spectrum by factors between 2 and 5
- Dirty lens 1: blends a single dirty lens pattern into an image with weights values between 0.2 and 0.65
- Dirty lens 2: overlays a distinct dirt pattern onto an image for each challenge level
- Salt & pepper noise: replaces random pixels of an image with either one or zero, with amount between 0.2 and 0.9
- Grayscale: convert an image into monochrome
To access the complete dataset and associated files, please refer to link.
Cite This Work:
@inproceedings{Temel2018_ICMLA,
author = {D. Temel and J. Lee and G. AlRegib},
booktitle = {2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA)},
title = {CURE-OR: Challenging unreal and real environments for object recognition},
year = {2018},}