HSD

This code is a official implementation of "Hierarchical Shot Detector (ICCV2019)" on COCO object detection with Pytorch.

Introduction

We propose a novel pipeline for accurate object detection (called ROC). Instead of simultaneous classification and regression, ROC firstly conducts box regression, secondly predicts the feature sampling locations for box classification, and finally classifies regressed boxes with the features of offset locations. To achieve the better detection accuracy, a hierarchical shot detector is proposed by stacking two ROC modules. Meanwhile, the contextual information is also incorporated to enrich the features of the second ROC module.

HSD detection pipeline.

Results

name	backbone	input size	minival	download
HSD320	VGG16	320x320	33.9	model
HSD512	VGG16	512x512	38.6	model
HSD512	ResNet101	512x512	39.7	model
HSD512	ResNext101	512x512	41.4	model

Installation

Install PyTorch-0.4.0 and Python 3+.
Clone this repository.
Compile the nms and install coco tools:

cd HSD
pip install Cython, pycocotools, opencv-python, matplotlib, PyYaml
./make.sh

Compile the deformable convolution:

cd HSD/models/deform
sh make.sh
CC=g++ python build.py

Then download the COCO dataset and put the MS COCO dataset at $path/data/coco

coco
|_ images
|  |_ train2014
|     |_ <imgname>.jpg
|     |_ ...
|  |_ val2014
|     |_ ...
|  |_ test2015
|     |_ ...
|_ annotations
|  |_ instances_train2014.json
|  |_ ...
|_ cache

Train and Inference

Download the pre-trained models (e.g., VGG16, Resnet, and ResNext) and put these models in the HSD/weights/pretrained_models/ dir:

VGG16:      wget https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth
ResNet101:  wget https://download.pytorch.org/models/resnet101-5d3b4d8f.pth
Resnext101: wget https://s3.ap-northeast-2.amazonaws.com/open-mmlab/pretrain/third_party/resnext101_32x4d-a5af3160.pth

Use train.py to train the detector:

e.g., python train.py --cfg ./configs/hsd_vgg_coco_320.yaml

To evaluate a trained network:

e.g., python eval.py --cfg ./configs/hsd_vgg_coco_320.yaml --weights ./weights/hsd-vgg320-coco/hsd_vgg_epoch_160_300.pth

Ciatation

If the project helps your research, please cite this paper.

@article{Cao_HSD_ICCV_2019,
  author =       {Jiale Cao and Yanwei Pang and Jungong Han and Xuelong Li},
  title =        {Hierarchical Shot Detector},
  journal =      {Proc. International Conference on Computer Vision},
  year =         {2019}
}

Acknowledgement

Many thanks to the open source codes, i.e., SSD_Pytorch, deformable-convolution-pytorch, mmdetection, and DANet.

fabiopoiesi / HSD

HSD