Improving Data-Scarce Image Classification Through Multimodal Synthetic Data Pretraining

Deep Learning algorithms and models greatly benefit from the release of large-scale datasets, also including synthetically generated data, when real-life data is scarce. Multimodal datasets feature more descriptive environmental information than single-sensor ones, but they are generally small and not widely accessible. In this paper, we construct a synthetically-generated image classification dataset consisting of grayscale camera images and depth information acquired from an 8x8-pixel Time-of-Flight sensor. We propose and evaluate six Convolutional Neural Network-based feature-level fusion models to integrate the multimodal data, outperforming the accuracy of the cameraonly model by up to 17% in real-world settings. By pretraining the model on synthetically-generated sample pairs, followed by fine-tuning it with only 16 real-domain samples, we outperform a non-pretrained counterpart by 35% while maintaining the storage constraints in the order of hundreds of kB. Our proposed convolutional model, pretrained on both synthetic and real-world sensor data, achieves a top-1 accuracy of 86.48%, proving the benefits of using multimodal datasets to train feature-level data fusion neural networks. Low-power emerging embedded microcontrollers, such as multi-core RISC-V systems-on-chip, are perfect candidates for running our model due to their reduced power consumption and parallel computing capabilities that speed up inference.