• currently avaliable for pedestratin detection(binary class), 2 domains.

• Using Faster RCNN as baseline

• Implement with Pytorch 0.2.0.

Follow faster-rcnn.pytorch

#Transfer learning:
__C.TRANSFER = True             #if this is False, it woule be baseline, Faster RCNN
__C.TRANSFER_SELECT = 'ALL'     #select "ALL", "POSITIVE", "BALANCE" or "RANDOM"
__C.TRANSFER_WEIGHT = 0.5       #alpha
__C.TRANSFER_GAMMA = 16         #gamma
__C.TRANSFER_LOSS_START = 6     #transfer loss start epoch
__C.TRANSFER_GRL = False        #use Grandient Reversal or not

Put your target data into Pascal Foramtted folder with folder name VOC2007 and link with data/VOCdevkit2007, put your source data into VOC2012 and link data/VOCdevkit2012 and follow the VOC_0712 training from py-faster-rcnn

To change other data format (e.g. Kitti) to Pascal_VOC format, please refer to Kitti2Pascal tutorial

Data: ECCV Workshop WIDER FACE AND PEDESTRIAN CHALLENGE 2018 Challenge 2 . Use Surveillance data set as source and Driven data set as target.