Faceboxes in Pytorch and trained with dataset FDDB and change the input size to 512*512
The original annotations of FDDB is ellipsel form, in order to be direct trained in Faceboxes we should change it to VOC form.
use the /data/FDDB/fddbannotation.py it will automaticlly generate the VOC annotations.
Also we should take care of the img_list.txt with the same form as WIDER_FACE(with both .jpg and .xml information). To transform the orginal img_list.txt using : data/FDDB/trans.py
change input size to 512*512 in the train.py
img_dim = 512 # only 1024 is supported
And train the FDDB as below:
python3 train.py --train './data/FDDB'
when we trained FACEBOXES with FDDB, then it could be test with WIDER_FACE.
As the same take care of the img_list.txt without xml info(otherwise the image can not be read). using /data/WIDER_FACE/trans.py