The Elephant in the Room
read the original abstract
We showcase a family of common failures of state-of-the art object detectors. These are obtained by replacing image sub-regions by another sub-image that contains a trained object. We call this "object transplanting". Modifying an image in this manner is shown to have a non-local impact on object detection. Slight changes in object position can affect its identity according to an object detector as well as that of other objects in the image. We provide some analysis and suggest possible reasons for the reported phenomena.
This paper has not been read by Pith yet.
Forward citations
Cited by 1 Pith paper
-
Metamorphic Testing of a Deep Learning based Forecaster
Developed 19 metamorphic relations to test correlation detection and LSTM forecasting in an outage prediction application, uncovering 8 unknown issues in the live system and detecting 65.9% of injected bugs via mutati...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.