hitdet.pytorch's Introduction

Hit-Detector Code Base

Implementation of our CVPR2020 paper Hit-Detector: Hierarchical Trinity Architecture Search for Object Detection

We released the searched Hit-Detector Architecture.

Environments

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.

Data Preparatoin

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
│       └── SegmentationObject

Getting Start

Our pretrained backbone params can be found in BaiduCloud. pwd: jbsm or GoogleDrive

Train the searched model:

cd scripts
sh train_hit_det.sh

Results on COCO minival

Model	Params	mAP
FPN	41.8M	36.6
Hit-Det	27.6M	41.3

Citation

@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}
}