Inferring relatedness from genomic data is an essential component of genetic association studies, population genetics, forensics, and genealogy. Due to the random nature of Mendelian inheritance, variance in the amount of the genome shared identically between two individuals of a certain degree of relatedness can be high, making relatedness inference difficult. While numerous methods exist for performing such inference, thorough evaluation of these methods in real data has been lacking. We assessed 11 state-of-the-art relatedness inference methods using a dataset with 2,485 individuals contained in several large pedigrees that span up to six generations. Overall, the methods have high accuracy (93%-99%) when reporting first and second degree relationships, but less than 60% accuracy for fifth degree relationships. We considered a composite method built off the three methods with highest accuracy in our analysis (ERSA 2.0, IBDseq, and Refined IBD) and applied it to the SAMAFS, HapMap3, and Weill Cornell Qatari datasets, finding numerous unreported relationships in all three datasets. Building on the insights from our analysis of methods, we developed DRUID---Deep Relatedness Utilizing Identity by Descent---a method that works by inferring the identical by descent (IBD) sharing profile of an ungenotyped ancestor of a set of close relatives. DRUID combines relatedness signals among multiple samples to effectively remove one or more generations of distance between a set of relatives, leading to substantial accuracy improvements compared to other methods.