Fiber optic shape sensors have enabled unique advances in various navigation tasks, from medical tool tracking to industrial applications. Eccentric fiber Bragg gratings (FBG) are cheap and easy-to-fabricate shape sensors that are often interrogated with simple setups. However, using low-cost interrogation systems for such intensity-based quasi-distributed sensors introduces further complications to the sensor's signal. Therefore, eccentric FBGs have not been able to accurately estimate complex multi-bend shapes. Here, we present a novel technique to overcome these limitations and provide accurate and precise shape estimation in eccentric FBG sensors. We investigate the most important bending-induced effects in curved optical fibers that are usually eliminated in intensity-based fiber sensors. These effects contain shape deformation information with a higher spatial resolution that we are now able to extract using deep learning techniques. We design a deep learning model based on a convolutional neural network that is trained to predict shapes given the sensor's spectra. We also provide a visual explanation, highlighting wavelength elements whose intensities are more relevant in making shape predictions. These findings imply that deep learning techniques benefit from the bending-induced effects that impact the desired signal in a complex manner. This is the first step toward cheap yet accurate fiber shape sensing solutions.