OVERVIEW
Global descriptors, localised.
SIMPLE began as a family of four descriptors: Simple-SCD, Simple-CLD, Simple-EHD and Local-CEDD, also known as LoCATe. Its central idea is to apply compact global descriptors to local image patches, allowing retrieval systems to describe salient regions rather than relying only on the image as a whole.
The initial scheme used the SURF detector to identify regions of interest. Instead of describing those regions with SURF, it applied MPEG-7 SCD, CLD and EHD, or the CEDD descriptor. The local features were then evaluated within a Bag of Visual Words framework for content-based image retrieval.
Generate image patches
SURF, SIFT or random patch generators identify the regions to be described.
Apply a visual descriptor
SCD, CLD, EHD, CEDD or CoMo describes the visual information within each patch.
Build the representation
A Bag of Visual Words model converts the local features into a retrieval-ready representation.
RESEARCH RATIONALE
From the whole image to its salient regions.
Early image-retrieval systems frequently represented visual content with a single global description. Evaluation on complex scenes and partially occluded objects showed that correctly ranked results can depend more strongly on salient regions than on the overall depiction. Representing an image through points and regions of interest therefore provides a more robust basis for many retrieval tasks.
SIMPLE descriptors focus on the characteristics required for an abstract but retrieval-friendly description of image patches. The framework was evaluated on established benchmarking databases.

OPEN SOURCE · GNU GPL
Implementations and downloads.
Official SIMPLE implementations are available in Java, C# and MATLAB. The Random Image Patches Generator is the suggested configuration.
RECOMMENDED
SIMPLE with Random Image Patches Generator.
The preferred implementation is available as Java source through LIRE and as a standalone C# package with source code and DLL.
ROBOTICS APPLICATION
LoCATe-based visual place recognition.
This visual place-recognition approach uses a visual vocabulary built from the Local CEDD descriptor to address loop-closure detection. Experimental results showed high recognition accuracy, while parallelisation supported real-time execution even on a low-cost mobile device.

CBMI 2014
The first SIMPLE presentation.
Slides from the original presentation of the SIMPLE descriptors at CBMI 2014.

