Based on jwyang/fpn.pytorch, i change little code to get a more reasonable mAP when training pascal voc 2007 and 07+12. Pytorch implementation of Feature Pyramid Network (FPN) for Object Detection.
This project inherits the property of our jwyang/fpn.pytorch.Hence, you can see more information about it.The following things are what I did :
- The stride of Resnet layer4 change 2 from 1. The most fundamental reason why mAP is low is that the anchor's position and number of each layer are calculated by stride in this code.The designed FPN_FEAT_STRIDES in config is [4, 8, 16, 32, 64]. When layer4's stride is set to 1, FPN_FEAT_STRIDES should be changed to [4, 8, 16, 16, 32], but FPN_FEAT_STRIDES is still the default value, which results in p5, p6 has about 3/4 of the anchors generated outside the image.
- Changing loge to log2 in _PyramidRoI_Feat.In original paper, roi pool on pyramid feature maps using log2. It does not seem to affect the training results.
- It supports training VOC07+12.In the original code, in order to batch training and memory efficient, it crop the original image.When i train VOC07+12, i find some images don't have target object duo to the operation of crop. So i add a paramter ASPECT_CROPPING in config.py, set it False , it will not crop the images. So you can train VOC07 + 12.
- It supports both python2 and python3.
I benchmark this code thoroughly on pascal voc2007 and 07+12. Below are the results:
1). PASCAL VOC 2007 (Train/Test: 07trainval/07test, scale=600, ROI Align，
| model | GPUs | Batch Size | lr | lr_decay | max_epoch | Speed/epoch | Memory/GPU | mAP |
|---|---|---|---|---|---|---|---|---|
| Res-101 | 1 GTX 1080 (Ti) | 2 | 1e-3 | 10 | 12 | 0.22 hr | 6137MB | 75.7 |
2). PASCAL VOC 07+12 (Train/Test: 07+12trainval/07test, scale=600, ROI Align)
| model | GPUs | Batch Size | lr | lr_decay | max_epoch | Speed/epoch | Memory/GPU | mAP |
|---|---|---|---|---|---|---|---|---|
| Res-101 | 1 GTX 1080 (Ti) | 1 | 1e-3 | 10 | 12 | \ | 9011MB | 80.5 |
First of all, clone the code
git clone https://github.com/guoruoqian/FPN_Pytorch.git
Then, create a folder:
cd FPN_Pytorch && mkdir data
- Python 2.7 or 3.6
- Pytorch 0.2.0 or higher
- CUDA 8.0 or higher
- tensorboardX
- VOC2007: Please follow the instructions in py-faster-rcnn to prepare VOC datasets. Actually, you can refer to any others. After downloading the data, creat softlinks in the folder data/.
- VOC 07 + 12: Please follow the instructions in YuwenXiong/py-R-FCN . I think this instruction is more helpful to prepare VOC datasets.
Please follow the instructions in Pretrained Model and Compilation.
train voc2007:
CUDA_VISIBLE_DEVICES=3 python3 trainval_net.py exp_name --dataset pascal_voc --net res101 --bs 2 --nw 4 --lr 1e-3 --epochs 12 --save_dir weights --cuda --use_tfboard True
test voc2007:
CUDA_VISIBLE_DEVICES=3 python3 test_net.py exp_name --dataset pascal_voc --net res101 --checksession 1 --checkepoch 7 --checkpoint 5010 --cuda --load_dir weights
train voc07+12:
CUDA_VISIBLE_DEVICES=3 python3 trainval_net.py exp_name2 --dataset pascal_voc_0712 --net res101 --bs 2 --nw 4 --lr 1e-3 --epochs 12 --save_dir weights --cuda --use_tfboard True