HrNet backbone in CenterNet

This code is use to train and evaluate the hornet backbone in CenterNet. For more technical details, please refer to the arXiv paper.

CenterNet is an one-stage detector which gets trained from scratch. On the MS-COCO dataset, CenterNet achieves an AP of 47.0%, which surpasses all known one-stage detectors, and even gets very close to the top-performance two-stage detectors.

Architecture

Preparation

Please first install Anaconda and create an Anaconda environment using the provided package list.

conda create --name CenterNet --file conda_packagelist.txt

After you create the environment, activate it.

source activate CenterNet

Compiling Corner Pooling Layers

cd <CenterNet dir>/models/py_utils/_cpools/
python setup.py install --user

Compiling NMS

cd <CenterNet dir>/external
make

Installing MS COCO APIs

cd <CenterNet dir>/data/coco/PythonAPI
make

Downloading MS COCO Data

• Download the training/validation split we use in our paper from here (originally from Faster R-CNN)
• Unzip the file and place annotations under <CenterNet dir>/data/coco
• Download the images (2014 Train, 2014 Val, 2017 Test) from here
• Create 3 directories, trainval2014, minival2014 and testdev2017, under <CenterNet dir>/data/coco/images/
• Copy the training/validation/testing images to the corresponding directories according to the annotation files

Training

To train HrCenterNet:

python train.py HRNet

We provide the configuration file (HRNet.json) and the model file (HRNet-104.py) for CenterNet in this repo.

To continue training:

1. modify the pretrained in HRNet.json
2. python train.py HRNet --iter 10000

To train DLANet:

python train.py DLANet

Evaluation

To test HRNet:

python test.py HRNet --testiter 480000 --split validation

To test DLANet:

python test.py DLANet --testiter <iter> --split <split>