This article is a "how to detect the segmentation result of FP and FN" were discussed
Abstract：
In recent years, image recognition direction, the depth of learning has gone beyond the other methods. Assist decision-makers from human to change more and more automated systems, increasing the demand failure modes correctly handle the depth learning modules. The contribution of this article, we review a number of techniques used to quantify the uncertainty in machine learning algorithms based on self-monitoring. In particular, we will apply semantic segmentation tasks, machine learning algorithm based on semantic decomposition image category. We discussed at the instance level of false positive and false negative error pattern, and recalled the detection of such errors technology recently proposed, we have for the future research directions are suggested.
Terms Index: Deep Learning, Semantic Segmentation, false positive and false negative Detection
一、Introduction
Depth learning techniques amazing achievement, especially CNN technology, technology development boom has led to new applications, it is unrealistic to believe that a decade ago. In particular, in the automotive industry, fully automated driving system has been widely used. Although such systems from the industry's efforts to assist the driver to drive to a higher level, this higher level can temporarily or completely replace the technology. How deep the problem based learning techniques designed autopilot system there are still many unresolved issues, especially in terms of reliability and security. When the AI-driven systems assist interpretation of medical images, there is a similar problem, although no intention of making the process fully automated.
We will focus on the semantic interpretation based on the road scene camera data, which is an important prerequisite for any autopilot strategy. To be more specific, we will focus on semantic segmentation, rather than target detection. Semantic segmentation, an image is decomposed into a plurality of masks, each mask pixel attached to a unified particular category in a predefined semantic space. Although the existence of each instance dividing network connected component, but here we considered an example of a mask. Based on these examples, the following failure modes have to be considered:
- False positive (FP)
- False negative (FN)
- Out of Distribution (OOD): Objects outside the semantic space, in which trained perception algorithm, but still appears in the input data, and therefore are misclassified
- Adversarial Attack (AA): a sensing module by manipulation of the input sensor, to some extent, forced to commit errors FP or FN
We focus on the first two "FP" and "FN" failure mode . In particular, we discuss the method of self-monitoring division of the network. We are not considered here due to the structure autonomous vehicles redundancy and increased reliability.
