vov-net / VoVNet-FCOS

FCOS with VoVNet(CVPRW'19) backbone networks

FCOS with VoVNet(CVPRW'19) Efficient Backbone Networks

This repository contains FCOS(ICCV'19) with VoVNet (CVPRW'19) efficient backbone networks. This code based on pytorch imeplementation of FCOS

Highlights

• Memory efficient
• Better performance, especially for small object
• Faster speed

Comparison with ResNet backbones

• same hyperparameters
• same training protocols ( max epoch, learning rate schedule, etc)
• 8 x TITAN Xp GPU
• pytorch1.1
• CUDA v9
• cuDNN v7.2

Backbone	Multi-scale training	Inference time (ms)	Box AP (AP/APs/APm/APl)	DOWNLOAD
R-50-FPN-1x	No	84	37.5/21.3/40.3/49.5	-
V-39-FPN-1x	No	82	37.7/22.4/41.8/48.4	link
R-101-FPN-2x	Yes	104	41.3/25.0/45.5/53.0	-
V-57-FPN-2x	Yes	91	41.6/25.9/45.6/53.1	link
R-101-32x8d-FPN-2x	Yes	171	42.5/26.0/46.1/54.2	-
V-93-FPN-2x	Yes	113	42.1/26.2/46.0/53.9	link

ImageNet pretrained weight

• VoVNet-39
• VoVNet-57
• VoVNet-75
• VoVNet-93

Installation

Check INSTALL.md for installation instructions which is orginate from FCOS

Training

Follow the instructions

For example,

# specify the number of GPU you can use.
export NGPUS=8 
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file "configs/vovnet/fcos_V_39_FPN_1x.yaml" 

Evaluation

Follow the instructions

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/8n0wyypfggliplw/FCOS-V-39-FPN-1x.pth?dl=1
export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/test_net.py --config-file "configs/vovnet/fcos_V_39_FPN_1x.yaml" TEST.IMS_PER_BATCH 16 MODEL.WEIGHT FCOS-V-39-FPN-1x.pth

single-gpu evaluation & test batch size 1,

wget https://dl.dropbox.com/s/8n0wyypfggliplw/FCOS-V-39-FPN-1x.pth?dl=1
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 FCOS-V-39-FPN-1x.pth

Related projects

• VoVNet-Classification
• VoVNet-Detectron