Can state-of -the-art visual analysis software always be relied on to find the elephant in an image? Akshay Raj Dhamija and colleagues at the VAST lab (Vision And Security Technology) at the Department of Computer Science, at University of Colorado, at Colorado Springs, US. have recently investigated the robustness of automated elephant detector systems.
The team point out that whilst object detection research has a long history in computer vision (more than 50 years) it’s not always 100% accurate. Even using the latest versions of large-scale deep-learning neural-networks.
“While current state-of-the-art detectors are trained to handle backgrounds, their designs are not well equipped to address unknown objects, which are often incorrectly detected as the existing classes with a high conﬁdence.”
[ note mis-identified elephants above ]
A recurring problem is that current systems still come across difficulties in deciding which regions in an image are ‘objects’ (of interest) and which are just insignificant backgrounds.
“Therefore, we consider it important that detection systems eventually learn to create a separation between background and unknown objects, enabling new objects to be identiﬁed. Currently, there is no such architecture and a design is left for the future work.”
See: Identifying the Elephant in Object Detection
Also see: (from the VAST lab) Detecting and Classifying Scars, Marks, and Tattoos Found in the Wild
[ research research by Martin Gardiner ]