Awesome
CornerNet
Reproduce of Cornernet
The original pytorch implementation repository is here
Requirements
-
You will need python modules: cv2, matplotlib and numpy.
-
To compile corner pooling layer, yo need to install mxnet 1.3.0, then put the files in cxx_operator into src/operator/nn/ in mxnet source code and compile it, then run
cd ${YOUR_MXNET_ROOT}
export PYTHONPATH=$(pwd)/lib/libmxnet.so:${PYTHONPATH}
to make sure you import the correct mxnet library.
Alternatively, you can uncomment line 92 and 93 and comment line 94, 95 in symbols/cornernet.py to use python implementation of cornerpooling layer, which would be much slower.
- run init.sh to compile nms and pycocotools
Demo results
mAP
| Model | Training data | Test data | mAP |
|---|---|---|---|
| CornerNet_coco_511x511 | train2014+valminusminival2014 | minival2014 | 38.9 |
TRAIN
You need to put the coco image files in date.
You can change the batch_size in config/cfg.py according to your gpu number and their computation abilies, but make sure that batch_size number is proportional to the number of gpus.
python train.py --gpus 0,1
TEST
Download the compressed model from CornerNet_coco_511x511 and unzip it then put it in model/, then run
python test.py --prefix model/cornernet --epoch 100 --gpus 0
if you want to visualize the test results:
python test.py --prefix model/cornernet --epoch 100 --gpus 0 --debug True
images will be saved in images/