SSD coco dataset 설정

https://github.com/weiliu89/caffe/tree/ssd/data/coco

1. Download Images and Annotations from MSCOCO. By default, we assume the data is stored in $HOME/data/coco


1. annotations 다운로드

MSCOCO 사이트에서 다운로드

instances_train-val2014.zip: instances_train2014.json, instances_val2014.json

image_info_test2014.zip: image_info_test2014.json

image_info_test2015.zip: image_info_test2015.json, image_info_test-dev2015.json

https://github.com/rbgirshick/py-faster-rcnn/tree/master/data 여기서 다운로드

https://dl.dropboxusercontent.com/s/s3tw5zcg7395368/instances_valminusminival2014.json.zip?dl=0 : instances_valminusminval2014.json.zip

https://dl.dropboxusercontent.com/s/o43o90bna78omob/instances_minival2014.json.zip?dl=0 : instances_minval2014.json.zip

2. image 다운로드

2014 Training images [80k/13GB]

2014 Val. images [40k/6.2GB]

2014 Testing images [40k/6.2GB]

2015 Testing images [80k/12.4G]

3. 디렉토리 설정

$HOME/data/coco 디렉토리 생성

~/coco/annotations 디렉토리 생성 + annotations 파일 옮기기

~/coco/images 디렉토리 생성 + images 파일 옮기기

2. Get the coco code. We will call the directory that you cloned coco into $COCO_ROOT

git clone https://github.com/weiliu89/coco.git
cd coco
git checkout dev

3. Build the coco code.

cd PythonAPI
python setup.py build_ext --inplace

4. Split the annotation to many files per image and get the image size info.

# Check scripts/batch_split_annotation.py and change settings accordingly.
python scripts/batch_split_annotation.py
# Create the minival2014_name_size.txt and test-dev2015_name_size.txt in $CAFFE_ROOT/data/coco
python scripts/batch_get_image_size.py

batch_get_image_size.py 파일에 $CAFFE_ROOT가 $HOME/projects/caffe로 되어있어서 새로 폴더를 만들어서 결과물을 여기에 저장함

자신의 caffe 경로로 옮겨줄것

5. Create the LMDB file.

cd $CAFFE_ROOT
# Create the minival.txt, testdev.txt, test.txt, train.txt in data/coco/
python data/coco/create_list.py
# You can modify the parameters in create_data.sh if needed.
# It will create lmdb files for minival, testdev, test, and train with encoded original image:
#   - $HOME/data/coco/lmdb/coco_minival_lmdb
#   - $HOME/data/coco/lmdb/coco_testdev_lmdb
#   - $HOME/data/coco/lmdb/coco_test_lmdb
#   - $HOME/data/coco/lmdb/coco_train_lmdb
# and make soft links at examples/coco/
./data/coco/create_data.sh

6. python examples/ssd/ssd_coco.py

gpu 수 알맞게 고치기 (line 334)

ex) gpus="0,1,2,3" -> gpus="0"