Abstract

Deep neural networks achieve remarkable performance in many computer vision tasks. For most networks in semantic segmentation (seg) and object detection (det) tasks, the backbone directly reuses the network manually designed for classification tasks. Utilizing a network pre-trained on ImageNet as the backbone has been a popular practice for seg/det tasks. However, because of the gaps between different tasks, adapting the network to the target task brings performance promotion. Some recent neural architecture search (NAS) methods search for the backbone of seg/det networks. ImageNet pre-training of the search space representation or the searched network bears huge computational cost. In this paper, we propose a fast neural network adaptation (FNA) method, which can adapt both the architecture and parameters of a manually designed seed network to the new seg/det tasks efficiently. A parameter remapping mechanism is designed to accelerate the whole adaptation process which takes full advantage of the knowledge from the seed network. The NAS method is utilized for architecture adaptation. In experiments, we conduct network adaptation on the MobileNetV2 network to both seg and det tasks. FNA demonstrates clear performance gains compared with both manually and NAS designed networks. The total computational cost of FNA is significantly less than many SOTA seg/det NAS methods: 1737× less than DPC, 6.8× less than Auto-DeepLab and 7.4× less than DetNAS.