Graphics Processing Units (GPUs) are currently the dominating programmable architecture for Deep Learning (DL) accelerators. The adoption of Field Programmable Gate Arrays (FPGAs) in DL accelerators is however getting momentum. In this paper, we demonstrate that Direct Hardware Mapping (DHM) of a Convolutional Neural Network (CNN) on an embedded FPGA substantially outperforms a GPU implementation in terms of energy efficiency and execution time. However, DHM is highly resource intensive and cannot fully substitute the GPU when implementing a state-of-the-art CNN. We thus propose a hybrid FPGA-GPU DL acceleration method and demonstrate that heterogeneous acceleration outperforms GPU acceleration even including communication overheads. Experimental results are conducted on a heterogeneous multi-platform setup embedding an Nvidia(R) Jetson TX2 CPU-GPU board and an Intel(R) Cyclone10GX FPGA board. The SqueezeNet, MobileNetv2, and ShuffleNetv2 mobile-oriented CNNs are experimented. We show that heterogeneous FPG-AGPU acceleration outperforms GPU acceleration for classification inference task over MobileNetv2 (12%-30% energy reduction, 4% to 26% latency reduction), SqueezeNet (21%-28% energy reduction, same latency), and ShuffleNetv2 (25% energy reduction, 21% latency reduction).