More and more video and image data is available to security authorities that can help solve crimes. Since manual analysis is time-consuming, algorithms are needed that support e.g. re-identification of persons. However, person re-identification approaches solely output image rank lists but do not provide an explanation for the results.
In this work, two concepts are proposed to explain person re-identification rankings and a qualitative evaluation is conducted. Both approaches are based on a multi-task convolutional neural network which outputs feature vectors for person re-identification and simultaneously recognizes a person’s semantic attributes. Analyses of the learned weights and the outputs of the attribute classifier are used to generate the explanations.
The results of the conducted experiments indicate that both approaches are suitable to improve the comprehensibility of person re-identification rankings.