Towards lightweight convolutional neural networks for object detection

摘要

We propose model with larger spatial size of feature maps and evaluate it onobject detection task. With the goal to choose the best feature extractionnetwork for our model we compare several popular lightweight networks. Afterthat we conduct a set of experiments with channels reduction algorithms inorder to accelerate execution. Our vehicle detection models are accurate, fastand therefore suit for embedded visual applications. With only 1.5 GFLOPs ourbest model gives 93.39 AP on validation subset of challenging DETRAC dataset.The smallest of our models is the first to achieve real-time inference speed onCPU with reasonable accuracy drop to 91.43 AP.