Optimizing Deep Learning Models for Object Detection

Deep learning models for object detection have gotten larger and larger over the years, spanning from 3.9M trainable parameters for EfficientDet to 209M for the AmoebaNet-based NAS-FPN detector. Different strategies are currently being researched in order to improve the efficiency of deep learning models for object detection, one of which is running the training and inference of the neural network in low precision. Interesting results have been achieved by researchers, starting from the original paradigm of using operators and doing the necessary operations in IEEE single precision (FP32), to achieving similar accuracies of the models using custom minifloat formats (FP8). The results can be pushed even further by using genetic algorithms for hyperparameter tuning, in order to find specific hyperparameter for the FP8 version of the model. In this paper, we will present the results of our experiments utilizing YOLOv3 with hybrid floating-point format (HFP8). One of the experiments in this paper shows how our solution can be used for checking social distancing guidelines which is a very important topic in the current COVID-19 pandemic.