Figure 5
From: Mice use robust and common strategies to discriminate natural scenes
Blurring the images and other attempts to reduce the fit error (RMSE). Several approaches were investigated to determine whether they could reduce the error of the psychometric curve fit. (a) To determine whether high spatial frequency features of the images (which could be imperceptible to the mice, but still influence SSIM calculations) caused the relatively high root mean squared error (RMSE) values, the images were filtered (blurred) with Gaussian filters of a range of standard deviations (σ). The inset shows the filtered target image. (b) Overall, blurring the images only marginally decreased the RMSE, and aggressive blurring led to even higher RMSE. (c) Registration of images prior to SSIM calculations did not improve the psychometric curve fit. (d) An alternative metric, pixel-wise RMSE between distractors and a target image, also did not improve the RMSE of a fit curve. (e) Another alternative metric, pixel-wise cross-correlation between distractors and the target image, also failed to yield low-RMSE curve fits. (f) Finally, calculating the SSIM between distractors and a distractor for the training phase (the “anti-target”), also failed to yield a low RMSE. RMSEs for (c–f) was 0.089, 0.095, 0.11 and 0.10, respectively.
