Awesome
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.
| 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 |
| V-39 | Faster | 3.9 | 78 | 39.8/23.7/42.6/51.5 | - | link |
| R-101 | Faster | 4.7 | 97 | 39.6/22.8/43.2/51.9 | - | link |
| V-57 | Faster | 4.4 | 87 | 40.8/24.8/43.8/52.4 | - | link |
| V-75 | Faster | 5.3 | 96 | 41.2/24.1/44.3/53.0 | - | link |
| X-101-64x4d* | Faster | - | 41.3/-/-/- | - | - | |
| X-101-32x8d* | Faster | - | 40.6/-/-/- | - | - | |
| V-93 | Faster | 6.1 | 110 | 41.8/24.8/45.1/53.8 | - | link |
| R-50 | Mask | 3.6 | 83 | 38.6/22.1/41.3/51.4 | 34.9/16.0/37.3/52.2 | link |
| V-39 | Mask | 4 | 81 | 41.0/24.6/43.9/53.1 | 36.7/17.9/39.3/53.0 | 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-57 | Mask | 4.5 | 90 | 42.0/25.1/44.9/53.8 | 37.5/18.3/39.8/54.3 | link |
| V-75 | Mask | 5.4 | 106 | 42.4/26.0/45.5/54.8 | 37.7/18.9/40.4/54.9 | link |
| X-101-64x4d* | Faster | - | 42.2/-/-/- | 37.2/-/-/- | - | |
| X-101-32x8d* | Faster | - | 41.7/-/-/- | 36.9/-/-/- | - | |
| V-93 | Mask | 6.7 | 114 | 42.7/24.9/45.8/55.3 | 38.0/17.7/40.9/55.2 | link |
ImageNet pretrained weight
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=$NGPUS tools/train_net.py --config-file "configs/vovnet/e2e_faster_rcnn_V_39_FPN_2x.yaml"
Evaluation
Follow the instruction
First of all, you have to download the weight file you want to inference.
For examaple,
multi-gpu evaluation & test batch size 16,
wget https://dl.dropbox.com/s/f1per2rj4pi8t71/FRCN-V-39-FPN-2x-norm.pth
export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/test_net.py --config-file "configs/vovnet/e2e_faster_rcnn_V_39_FPN_2x.yaml" TEST.IMS_PER_BATCH 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