Implementation of our CVPR2020 paper Hit-Detector: Hierarchical Trinity Architecture Search for Object Detection
We released the searched Hit-Detector Architecture.
- Python 3.6
- Pytorch>=1.1.0
- Torchvision == 0.3.0
You can directly run the code sh env.sh to setup the running environment.
We use 8 GPUs (32GB V100) to train our detector, you can adjust the batch size in configs by yourselves.
Your directory tree should be look like this:
$HitDet.pytorch/data
├── coco
│ ├── annotations
│ ├── train2017
│ └── val2017
│
├── VOCdevkit
│ ├── VOC2007
│ │ ├── Annotations
│ │ ├── ImageSets
│ │ ├── JPEGImages
│ │ ├── SegmentationClass
│ │ └── SegmentationObject
│ └── VOC2012
│ ├── Annotations
│ ├── ImageSets
│ ├── JPEGImages
│ ├── SegmentationClass
│ └── SegmentationObjectOur pretrained backbone params can be found in BaiduCloud. pwd: jbsm or GoogleDrive
Train the searched model:
cd scripts
sh train_hit_det.sh
| Model | Params | mAP |
|---|---|---|
| FPN | 41.8M | 36.6 |
| Hit-Det | 27.6M | 41.3 |
@InProceedings{guo2020hit,
author = {Guo, Jianyuan and Han, Kai and Wang, Yunhe and Zhang, Chao and Yang, Zhaohui and Wu, Han and Chen, Xinghao and Xu, Chang},
title = {Hit-Detector: Hierarchical Trinity Architecture Search for Object Detection},
booktitle = {arXiv preprint arXiv:2003.11818},
year = {2020}
}
Our code is based on the open source project MMDetection.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent