ERF-NAS: Efficient Receptive Field-Based Zero-Shot NAS for Object Detection

Abstract

Neural Architecture Search (NAS) is widely used to design efficient and high-performance architectures for various models. However, most NAS methods for object detection tasks are complex and time-consuming, requiring substantial computational resources to train numerous candidate detectors in a vast search space. To address this issue, we propose ERF-NAS, an efficient receptive field-based zero-shot NAS method. First, we utilize the Effective Receptive Field (ERF) of the detector’s backbone network as a zero-cost proxy to assess the candidate architecture’s feature extraction quality and expression capabilities. The calculation of ERF only requires a single inference forward of a randomly initialized backbone network, eliminating the necessity for training. Second, to obtain better coordination between different network components, we introduce the transformation paradigm to adjust the depth of the neck network. To provide a more effective and accurate latency constraint of candidate architectures, we construct a latency table by measuring convolution operations in TensorRT format. Extensive experiments on MS COCO dataset show that ERF-NAS achieves superior accuracy-efficiency trade-off results with less than 1 GPU day search costs.