DETNAS: BACKBONE SEARCH FOR OBJECT DETECTION

Abstract

Object detectors are usually equipped with backbone networks designed for image classification. It might be sub-optimal because of the gap between the tasks of image classification and object detection. In this work, we present DetNAS to use Neural Architecture Search (NAS) for the design of better backbones for object detection. It is non-trivial because detection training typically needs ImageNet pre-training while NAS systems require accuracies on the target detection task as supervisory signals. Based on the technique of one-shot supernet, which contains all possible networks in the search space, we propose a framework for backbone search on object detection. We train the supernet under the typical detector training schedule: ImageNet pre-training and detection fine-tuning. Then, the architecture search is performed on the trained supernet, using the detection task as the guidance. This framework makes NAS on backbones very efficient. In experiments, we show the effectiveness of DetNAS on various detectors, for instance, one-stage RetinaNet and the two-stage FPN. We empirically find that networks searched on object detection shows consistent superiority compared to those searched on ImageNet classification. The resulting architecture achieves superior performance than hand-crafted networks on COCO with much less FLOPs complexity.

Citation

@article{chen2019detnas,
  title={Detnas: Backbone search for object detection},
  author={Chen, Yukang and Yang, Tong and Zhang, Xiangyu and Meng, Gaofeng and Xiao, Xinyu and Sun, Jian},
  journal={Advances in Neural Information Processing Systems},
  volume={32},
  pages={6642--6652},
  year={2019}
}

Results and models

Note:

1. The experiment settings of DetNAS are similar with SPOS's, and our training dataset is COCO2017 rather than COCO2014.
2. We also retrained official subnet with same experiment settings, the final result is 36.9

Getting Started

Step 1: Supernet pre-training on ImageNet

python ./tools/mmdet/train_mmdet.py \
  configs/nas/detnas/detnas_shufflenet_supernet_imagenet.py \
  --work-dir $WORK_DIR

Step 2: Supernet fine-tuning on COCO

python ./tools/mmdet/train_mmdet.py \
  configs/nas/detnas/detnas_frcnn_shufflenet_fpn_supernet_coco_1x.py \
  --work-dir $WORK_DIR \
  --cfg-options load_from=$STEP1_CKPT

Step 3: Search for subnet on the trained supernet

python ./tools/mmdet/search_mmdet.py \
  configs/nas/detnas/detnas_frcnn_shufflenet_fpn_evolution_search_coco.py \
  $STEP2_CKPT \
  --work-dir $WORK_DIR

Step 4: Subnet retraining on ImageNet

python ./tools/mmdet/train_mmdet.py \
  configs/nas/detnas/detnas_shufflenet_subnet_imagenet.py \
  --work-dir $WORK_DIR \
  --cfg-options algorithm.mutable_cfg=$STEP3_SUBNET_YAML

Step 5: Subnet fine-tuning on COCO

python ./tools/mmdet/train_mmdet.py \
  configs/nas/detnas/detnas_frcnn_shufflenet_fpn_subnet_coco_1x.py \
  --work-dir $WORK_DIR \
  --cfg-options algorithm.mutable_cfg=$STEP3_SUBNET_YAML load_from=$STEP$_CKPT