Response index: quantitative evaluation index of translational equivariance

Translational equivariance, one of the properties of Convolutional neural networks(CNNs), directly reflects the coherence of the influence of input at each position on the output. By looking for changes in variability such as translational equivariance, it is possible to determine whether the direction of model fit is correct. A controllable location target is designed to verify the translationlal equivariance of a CNN and then the effect of the CNN’s parameters on positioning errors was investigated. Furthermore, A quantitative method called response index(ResIndex) is proposed in this paper. When the parameters of a CNN are determined, the distribution of the input signal response at each position in the heatmap can be obtained via simple algebraic calculations. Here we demonstrate that translational equivariance is primarily affected by the convolution boundary effect,which can be quantitatively assessed by the ResIndex. Experimental evidence for the Pearson correlation coefficient between the MSE and ResIndex demonstrates that our ResIndex is strongly negatively correlated with the MSE, with the mean Pearson correlation coefficient is -0.9282 on the CIFAR-10 and -0.7837 on COCO. For the first time, a unified quantitative evaluation index called the ResIndex is proposed to measure the translational equivariance of CNN. A complete mathematical derivation and a time-saving calculation method are given.