Detectron with VoVNet(CVPRW'19) Backbone Networks

This repository contains Detectron with VoVNet (CVPRW'19) Backbone Networks. This code based on pytorch imeplementation of Detectron (maskrcnn-benchmark)

Highlights

• Memory efficient

• Better performance, especially for small objects

• Faster speed

Comparison with ResNet backbones

• 2x schedule

• same hyperparameters

• same training protocols ( max epoch, learning rate schedule, etc)

• NOT multi-scale training augmentation

• 8 x TITAN Xp GPU

• pytorch1.1

• CUDA v9

• cuDNN v7.2

Note

'*' indicates ResNeXt models from Detectron official caffe2 page.

ImageNet pretrained weight

• VoVNet-39

• VoVNet-57

• VoVNet-75

• VoVNet-93

Installation

Check INSTALL.md for installation instructions which is orginate from maskrcnn-benchmark

Training

Follow the instructions maskrcnn-benchmark guides

For example,

export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file "configs/vovnet/e2e_faster_rcnn_V_39_FPN_2x.yaml"

Evaluation

Follow the instruction

First of all, you have to download the weight file you want to inference.

For examaple,

multi-gpu evaluation & test batch size 16,

wget https://dl.dropbox.com/s/f1per2rj4pi8t71/FRCN-V-39-FPN-2x-norm.pthexport NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/test_net.py --config-file "configs/vovnet/e2e_faster_rcnn_V_39_FPN_2x.yaml"   TEST.IMS_PER_BATCH 16 MODEL.WEIGHT FRCN-V-39-FPN-2x-norm.pth

For single-gpu evaluation & test batch size 1,

wget https://dl.dropbox.com/s/f1per2rj4pi8t71/FRCN-V-39-FPN-2x-norm.pth
CUDA_VISIBLE_DEVICES=0
python tools/test_net.py --config-file "configs/vovnet/e2e_faster_rcnn_V_39_FPN_2x.yaml" TEST.IMS_PER_BATCH 1 MODEL.WEIGHT FRCN-V-39-FPN-2x-norm.pth

TO DO LIST

•  super slim models

•  slim models

•  Larger models

•  Multi-scale training & test