Spatial Recall Index for the performance of machine-learning algorithms for automotive applications
We simulate a realistic objective lens based on a Cooke-triplet that exhibits typical optical aberrations like astigmatism and chromatic aberration, all variable over field. We use a special pixel-based convolution to degrade a subset of images from the BDD100k dataset, and quantify the changes in the performance of the pre-trained Hybrid Task Cascade (HTC) and Mask R-CNN algorithm. We present the SRI, which spatially resolves where in the image these changes occur, on a pixel-by-pixel basis. Our examples demonstrate the spatial dependence of the performance from the optical quality over field, highlighting the need to take the spatial dimension into account when training ML-based algorithms, especially when looking forward to autonomous driving applications.