Home

Awesome

CornerNet

Reproduce of Cornernet

The original pytorch implementation repository is here

Requirements

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.

Demo results

demo1

demo2

mAP

ModelTraining dataTest datamAP
CornerNet_coco_511x511train2014+valminusminival2014minival201438.9

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/