This repository contains Detectron with VoVNet (CVPRW'19) Backbone Networks. This code based on pytorch imeplementation of Detectron (maskrcnn-benchmark)
- Memory efficient
- Better performance, especially for small object
- Faster speed
- 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
| Backbone | Detector | Train mem(GB) | Inference time (ms) | Box AP (AP/APs/APm/APl) | Mask AP (AP/APs/APm/APl) | DOWNLOAD |
|---|---|---|---|---|---|---|
| R-50 | Faster | 3.6 | 78 | 37.5/21.3/40.3/49.5 | - | link |
| R-101 | Faster | 4.7 | 97 | 39.6/22.8/43.2/51.9 | - | link |
| V-39 | Faster | 3.9 | 78 | 39.8/23.7/42.6/51.5 | - | link |
| V-57 | Faster | 4.4 | 87 | 40.8/24.8/43.8/52.4 | - | link |
| R-50 | Mask | 3.6 | 83 | 38.6/22.1/41.3/51.4 | 34.9/16.0/37.3/52.2 | link |
| R-101 | Mask | 4.7 | 102 | 40.8/23.2/44.0/53.9 | 36.7/16.7/39.4/54.3 | link |
| V-39 | Mask | 4 | 81 | 41.0/24.6/43.9/53.1 | 36.7/17.9/39.3/53.0 | link |
| V-57 | Mask | 4.5 | 90 | 42.0/25.1/44.9/53.8 | 37.5/18.3/39.8/54.3 | link |
git clone https://github.com/stigma0617/maskrcnn-benchmark-vovnet.git
cd maskrcnn_benchmark-vovnet
git checkout vovnet
mkdir -p checkpoints/pretrained
cd checkpoints/pretrained
wget https://www.dropbox.com/s/b826phjle6kbamu/vovnet57_statedict_norm.pth
wget https://www.dropbox.com/s/s7f4vyfybyc9qpr/vovnet39_statedict_norm.pthCheck INSTALL.md for installation instructions which is orginate from maskrcnn-benchmark
Follow the instructions maskrcnn-benchmark guides
For example,
cd maskrcnn_benchmark-vovnet
export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file "configs/vovnet/e2e_faster_rcnn_V_39_FPN_2x.yaml" Follow the instruction
For multi-gpu evaluation & test batch size 16,
export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/test_net.py --config-file "configs/vovnet/e2e_faster_rcnn_V_39_FPN_2x.yaml" TEST.IMS_PER_BATCH 16For single-gpu evaluation & test batch size 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