{"title":"通过一次性神经结构搜索设计目标检测网络","authors":"Chuntung Zhuang","doi":"10.1109/ASSP54407.2021.00013","DOIUrl":null,"url":null,"abstract":"Previous NAS focus on image classification, so directly implementing traditional NAS, like the last research work DetNet, on object detection tasks are ineffective. In general, conventional NAS only searches the backbone network architecture while completely ignore the head network. Unlike image classification tasks, which can directly perform NAS on classification datasets such as ImageNet, object detection tasks require iterative training on classification and detection datasets multiple times. In addition, using traditional NAS methods on object detection tasks is computation-intensive (more than hundreds of GPU hours) because NAS typically has a two-stage workflow. The best sub-Network architecture derived from the super-Network must be retrained or fine-tuned. To resolve the above challenges, we propose DetNAS, a new NAS method targeting object detection tasks. First of all, DetNAS can accelerate the search process of neural network architecture to meet the various demands of object detection tasks. DetNet only searches the backbone network for object detection tasks, while DetNAS can simultaneously search for the backbone and head network during network architecture search. At the same time, inspired by the previous work, DetNAS replaced the single-path evolutionary algorithm in DetNet with progressive search, which further improved the search efficiency of the network structure. Secondly, previous research suggested a series of techniques to improve image classification-based network architecture search efficiency, but their effects on object detection tasks are still unknown. To thoroughly verify the validity of these conclusions, we conducted ablation experiments on multiple datasets and various experimental settings, which provided a valuable basis and reference for subsequent research work. As far as we know, DetNAS is the first One-shot NAS method that can search the backbone and head network simultaneously. We believe our work will open up a new direction to explore the architecture of object detection models.","PeriodicalId":153782,"journal":{"name":"2021 2nd Asia Symposium on Signal Processing (ASSP)","volume":"282 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"DetNAS: Design Object Detection Network via One-Shot Neural Architecture Search\",\"authors\":\"Chuntung Zhuang\",\"doi\":\"10.1109/ASSP54407.2021.00013\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"Previous NAS focus on image classification, so directly implementing traditional NAS, like the last research work DetNet, on object detection tasks are ineffective. In general, conventional NAS only searches the backbone network architecture while completely ignore the head network. Unlike image classification tasks, which can directly perform NAS on classification datasets such as ImageNet, object detection tasks require iterative training on classification and detection datasets multiple times. In addition, using traditional NAS methods on object detection tasks is computation-intensive (more than hundreds of GPU hours) because NAS typically has a two-stage workflow. The best sub-Network architecture derived from the super-Network must be retrained or fine-tuned. To resolve the above challenges, we propose DetNAS, a new NAS method targeting object detection tasks. First of all, DetNAS can accelerate the search process of neural network architecture to meet the various demands of object detection tasks. DetNet only searches the backbone network for object detection tasks, while DetNAS can simultaneously search for the backbone and head network during network architecture search. At the same time, inspired by the previous work, DetNAS replaced the single-path evolutionary algorithm in DetNet with progressive search, which further improved the search efficiency of the network structure. Secondly, previous research suggested a series of techniques to improve image classification-based network architecture search efficiency, but their effects on object detection tasks are still unknown. To thoroughly verify the validity of these conclusions, we conducted ablation experiments on multiple datasets and various experimental settings, which provided a valuable basis and reference for subsequent research work. As far as we know, DetNAS is the first One-shot NAS method that can search the backbone and head network simultaneously. We believe our work will open up a new direction to explore the architecture of object detection models.\",\"PeriodicalId\":153782,\"journal\":{\"name\":\"2021 2nd Asia Symposium on Signal Processing (ASSP)\",\"volume\":\"282 1\",\"pages\":\"0\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2021-11-01\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"2021 2nd Asia Symposium on Signal Processing (ASSP)\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.1109/ASSP54407.2021.00013\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"2021 2nd Asia Symposium on Signal Processing (ASSP)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/ASSP54407.2021.00013","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
DetNAS: Design Object Detection Network via One-Shot Neural Architecture Search
Previous NAS focus on image classification, so directly implementing traditional NAS, like the last research work DetNet, on object detection tasks are ineffective. In general, conventional NAS only searches the backbone network architecture while completely ignore the head network. Unlike image classification tasks, which can directly perform NAS on classification datasets such as ImageNet, object detection tasks require iterative training on classification and detection datasets multiple times. In addition, using traditional NAS methods on object detection tasks is computation-intensive (more than hundreds of GPU hours) because NAS typically has a two-stage workflow. The best sub-Network architecture derived from the super-Network must be retrained or fine-tuned. To resolve the above challenges, we propose DetNAS, a new NAS method targeting object detection tasks. First of all, DetNAS can accelerate the search process of neural network architecture to meet the various demands of object detection tasks. DetNet only searches the backbone network for object detection tasks, while DetNAS can simultaneously search for the backbone and head network during network architecture search. At the same time, inspired by the previous work, DetNAS replaced the single-path evolutionary algorithm in DetNet with progressive search, which further improved the search efficiency of the network structure. Secondly, previous research suggested a series of techniques to improve image classification-based network architecture search efficiency, but their effects on object detection tasks are still unknown. To thoroughly verify the validity of these conclusions, we conducted ablation experiments on multiple datasets and various experimental settings, which provided a valuable basis and reference for subsequent research work. As far as we know, DetNAS is the first One-shot NAS method that can search the backbone and head network simultaneously. We believe our work will open up a new direction to explore the architecture of object detection models.