Convolutional networks (CNNs) achieve state-of-the-art performance in object detection (OD), but require large annotated image sets for training. A recent methodology, named Feature Learning from Image Markers (FLIM), trains CNNs using user-drawn image-markers placed on very few training images. In FLIM, convolutional kernels are learned directly from image patches extracted from the markers. Recently FLIM encoders have been coupled with one-layer convolutional adaptive decoders to create end-to-end object detectors, where the features are linearly combined by weights computed on-the-fly without ground-truth. For such, features have to ideally isolate object and background activations, thus, a z-score normalization using the scribbles’ statistics, named marker-based normalization (MBN), has been crucial. With MBN, the user has to solve marker placement for learning the kernel coefficients and the normalization parameters, which is not an intuitive task. In this work, we detach the marker sets used for kernels and MBN parameter learning, investigate MBN’s impact without user bias, and propose guidelines and automatic marker extensions to make user interaction more intuitive. We propose a marker-bot that uses pixel-wise ground-truth to draw scribbles with different ratios of foreground and background pixels, which are then used to evaluate multiple FLIM CNNs using five quantitative metrics and feature projection analysis. Then, we propose an intuitive guideline of marker drawing for kernel learning, and a superpixel-based marker-extension method adapts these markers to MBN suitable ones. Our results suggest that detaching the marker sets is beneficial for FLIM networks, and that the superpixel-extended markers consistently improve the metrics given suitable hyper-parameters.