Skip to content

Latest commit

 

History

History
91 lines (63 loc) · 2.61 KB

README.md

File metadata and controls

91 lines (63 loc) · 2.61 KB
Raw

Simple-CenterNet

PyTorch Implementation of CenterNet(Object as Points)

  • You don't need to bulid some cpp code to use Deformable Convolution used in CenterNet.

Performance

On VOC(Training:0712 trainval, Test:07)

Repo Backbone 0.5 mAP Trained model
This Repo ResNet-18 78.1
xingyizhou/CenterNet ResNet-18 75.6
Ximilar-com/xcenternet ResNet-18 70.0
xuannianz/keras-CenterNet ResNet-50 72.9
bubbliiiing/centernet-keras ResNet-50 77.1

On COCO

Repo Backbone mAP Trained model
This Repo ResNet-18
xingyizhou/CenterNet ResNet-18 28.1

What's difference between paper and this repo?

VOC

Paper

Refer to Appendix D: Experiments on PascalVOC in the paper

  • Learning rate scheduler: MultiStepLR(milestones=[45, 60], gamma=0.1)
  • Augmentation: HorizontalFlip, RandomScale, RandomTranslation, RandomCrop, and Color Jittering
  • Kernel size of max pooling: 3

This Repo

  • Learning rate scheduler: CosineDecay(per iteration)
  • Augmentation: HorizontalFlip, RandomScale, RandomTranslation, RandomCrop, Mosaic, Mixup(with Mosaic + 1.0 AP), and Color Jittering
  • Kernel size of max pooling: 7
  • Gaussian Kernel Generation Method: followed the method proposed in Training-Time-Friendly Network for Real-Time Object Detection (It’s not carefully selected. I just think that it is more reasonable than original one.)

COCO17

Paper

This Repo

Setup

git clone https://github.com/developer0hye/Simple-CenterNet
cd Simple-CenterNet

if (your_os == 'Window'):

scripts/download-voc0712.bat
scripts/download-coco17.bat

else:

scripts/download-voc0712.sh
scripts/download-coco17.sh

Training

VOC07+12

python train.py --data ./data/voc0712.yaml --step-batch-size 32 --forward-batch-size 32 --total-epoch 70

If your gpu memory is too lower to train the model, you should try to reduce forward-batch-size.

COCO17

Evaluation

VOC07+12

python eval.py --data ./data/voc0712.yaml --weights your_model.pth --flip

COCO17

Things we tried that didn't work

  • Random Rotation Augmentation