The Ethical Dilemma When (Not) Setting up Cost-Based Decision Rules in Semantic Segmentation

Abstract

Neural networks for semantic segmentation can be seen as statistical models that provide for each pixel of one image a probability distribution on predefined classes. The predicted class is then usually obtained by the maximum aposteriori probability (MAP) which is known as Bayes rule in decision theory. From decision theory we also know that the Bayes rule is optimal regarding the simple symmetric cost function. Therefore, it weights each type of confusion between two different classes equally, e.g., given images of urban street scenes there is no distinction in the cost function if the network confuses a person with a street or a building with a tree. Intuitively, there might be confusions of classes that are more important to avoid than others. In this work, we want to raise awareness of the possibility of explicitly defining confusion costs and the associated ethical difficulties if it comes down to providing numbers. We define two cost functions from different extreme perspectives, an egoistic and an altruistic one, and show how safety relevant quantities like precision / recall and (segment-wise) false positive / negative rate change when interpolating between MAP, egoistic and altruistic decision rules.

Generalized Few-Shot Semantic Segmentation: All You Need is Fine-Tuning

“…Since the amount of possible triplets for each class is large, for latency reasons we only sample min(|T c |, τ ) triplets for each class. 2 Let C + be the set of all points of the same arbitrary class, and let C − be the set of all points from other classes such that C + ∩ C − = ∅. To construct T c , we first randomly sample τ points from C + for our anchor points, τ points from C + that are disjoint from our anchor points to act as positive examples, and τ points from C − for our negative points.…”

“…As discussed in Chan et al [2], the broader issue among all semantic segmentation algorithms is that decision boundaries must be made when performing classification. As a consequence of this, in some applications (e.g.…”

Detection of False Positive and False Negative Samples in Semantic Segmentation

“…(1). Although it seems reasonable, according to common human sense, to assume that ψ z (y , y) should be different depending on the type of confusion, another decision policy may reveal ethical problems when it comes down to providing explicit numbers [22]. Therefore, the choice of cost functions to increase the sensitivity towards rare objects is subjected to constraints.…”

“…As alternatives we propose a decision principle -the maximum likelihood (ML) decision rule [20] -that looks out for the best fit of the data to a given semantic class. We review the false negative detection [21] using the ML decision rule for semantic segmentation and also discuss cost based decision rules in general along with the problems of setting the cost structure up [22].…”

Software architecture for human- centered reliability assessment for neural networks in autonomous

“…an instance based assessment, like overlooked vulnerable road users (VRUs). Information on instances is often evident that a measurement of performance based on pixel coverage is insufficient and should be replaced by the class specific asymmetry of importance of confusion events (Chan et al, 2019;Chan et al, 2020), it is confused with a lamppost, or fatal, if a pedestrian is overlooked due to a confusion with the street. Apart from For the example of autonomous driving, errors in perception could either be irrelevant, like if a tree is account the application specific failure modes.…”