Awesome
| Installation | Documentation | Tutorials |
Cascade R-CNN
forked from gluon-cv
edited by Yuerong Chen and Ning Li
Benchmarking
- PASCAL VOC 2007 (Train/Test: 2007+2012trainval/2007test, shorter size: 600)
| model | #GPUs | bs | lr | wd | epoch | decay | AP50 | AP75 | AP | AP Reference |
|---|---|---|---|---|---|---|---|---|---|---|
| VGG-RPN-baseline(pruned) | 2 | 1 | 1e-3 | 5e-4 | 20 | 14 | 75 | wait | wait | 42.9 |
| VGG-RPN-cascade(pruned) | 2 | 1 | 1e-3 | 5e-4 | 20 | 14 | 76.20 | 57.79 | 53.05 | 51.2 |
| RESNET50-RPN-cascade | 8 | 1 | 4e-3 | 5e-4 | 20 | 14 | 79.71 | 60.00 | wait | None |
| Res50-RFCN-Cascade | 8 | 1 | 2e-3 | 1e-4 | 20 | 14 | 76.87 | 55.80 | 50.95 | 51.8 |
| Res101-RFCN-Cascade | 8 | 1 | 2e-3 | 1e-4 | 20 | 14 | 79.12 | 59.45 | 54.79 | 54.2 |
- MS-COCO (Train/Test: train2017/val2017, shorter size: 800 for FPN and 600 for the others)
| model | #GPUs | bs | epoch | decay | AP | AP Reference |
|---|---|---|---|---|---|---|
| Res101-RFCN-Cascade | 8 | 1 | 24 | 16 | 34.4 | 33.3 |
Developing Environment
MXNet 1.3.0
The repo is based on the 1.3.x version MXNet. You will probably need to go to the MXNet official github repo and compile it.
If your MXNet installed by using pip install(at least on August). It should be of MXNet version 1.2.x, which is too old for this repo. You might need to solve the related problems because of the version mismatch.
Installation
-
Clone the cascade_rcnn_gluon repository, and we'll call the directory that you cloned cascade_rcnn_gluon into CASCADE_ROOT
git clone https://github.com/zzdang/cascade_rcnn_gluon.git -
Build cascade_rcnn_gluon
cd $CASCADE_ROOT/ # Follow the gluon-cv installation instructions here: # https://gluon-cv.mxnet.io/ python setup.py install
Training Cascade-RCNN
-
Get the training data
# This will download the pascal_voc dataset python scripts/datasets/pascal_voc.py -
Download the pretrained models on ImageNet. For VGG-Net(we called vgg16_pruned), the FC layers are pruned and 2048 units per FC layer are remained. In addition, the two FC layers are copied three times for Cascade R-CNN training.
# Download pre-trained model(You can download it use dropbox or baiduyun link) -[dropbox link](https://www.dropbox.com/s/tjgcwqgber2tlxh/VGG_16_fc2048_prune.params?dl=0) -[baiduyun link](https://pan.baidu.com/s/1RgG33zy40ssdWHdhPx0-Kg) passwd: b7ev # copy the pre-trained models to $CASCADE_ROOT/models/ cp /PATH/TO/DOWNLOAD/MODEL $CASCADE_ROOT/models/# convert the vgg16_pruned pretained params to vgg 16_pruned_cascade params python load_params.py -
training for Cascade-RCNN
# training for pruned VGG16 python scripts/detection/cascade_rcnn/train_cascade_rcnn.py --network vgg16_pruned -
Testing Demo
# testing for pruned VGG16 (VOC dataset) python scripts/detection/cascade_rcnn/demo_cascade_rcnn.py --network cascade_rcnn_vgg16_pruned_voc --pretrained /PATH/TO/TRAINED/MODEL
Training Cascade-RFCN
-
training for Cascade-RFCN
# training for resnet101_v1b python scripts/detection/cascade_rcnn/train_cascade_rfcn.py --network resnet101_v1b --lr 0.002 --wd 0.0001 --save-prefix ./models/ --gpus 0
Examples
To Do List
- Train Cascade-RCNN with FPN backbone