Description

Estimate head pose with CNN,using backbone convolution network such as ResNet / MobileNet / ShuffleNet.Also try regress face orientation vector [x,y,z] directly and regress the Expectation of classify softmax results. What's more,build soft label for classify.

Usage

Training

python train.py --mode [reg,cls,sord] --net [resnet,shufflenet,mobilenet]
--num_classes [33,66] --num_epochs --lr --lr_decay --unfreeze 
--train_data --valid_data --input_size [224,196,160,128,96] 
--width_mult [1.0,0.5] --batch_size --top_k --cls2reg --alpha
--save_dir

• mode: method for predict last result,choice from [reg,cls,sord],reg means regress directly,cls means regress the expectation of classify softmax results,sord means using soft label for cls
• net: the alternative backbone network
• num_classes: numbers of classify when mode is cls or sord
• lr: learning rate
• lr_decay: decay rate of learning rate,means lr=lr*lr_decay at assigned time
• unfreeze: for fine tune pretrained weight when epoch < unfreeze
• train_data: path of training dataset
• valid_data: path of validation dataset
• input_size: size of input image for network
• width_mult: value of width_mult for mobilenet
• batch_size: training and validation batch size
• top_k: top k of classify sofmax result
• cls2reg: method for calculate regression value after classify
• alpha: weight of regression loss
• save_dir: path of saving model weight snapshot

Testing

python test.py --mode [reg,cls,sord] --net [resnet,shufflenet,mobilenet]
--num_classes [33,66] --test_data --input_size [224,196,160,128,96] 
--width_mult [1.0,0.5] --batch_size --top_k --cls2reg --alpha
--snapshot --show_front --analysis --huge_error --collect_score
--save_dir

• mode: method for predict last result,choice from [reg,cls,sord],reg means regress directly,cls means regress the expectation of classify softmax results,sord means using soft label for cls
• net: the alternative backbone network
• num_classes: numbers of classify when mode is cls or sord
• test_data: path of testing dataset
• input_size: size of input image for network
• width_mult: value of width_mult for MobileNet
• batch_size: training and validation batch size
• top_k: top k of classify softmax result
• cls2reg: method for calculate regression value after classify
• alpha: weight of regression loss
• snapshot: path of model weight
• show_front: draw gt axis and predicted vector in originaa images and save
• analysis: analysis degrees distribution on euler angles and save analysis results
• huge_error: choice worst degrees predicted images and save
• collect_score: dump degrees error on every image into a json file

Experiment

Training Dataset：

	Blur(P)	gary(P)	amount	type
300W_LP	0.05	0.85	126k	color
DMS591	0.05	0	19K	infrared

Experiment expression: (degrees mae/collect score(limit=10º))

	pretrained	width_mult	input_size	Flops	Test on AFLW	Test on DMS
MobileNetV2	Yes	1.0	224	299	5.9º/88%	5.5º/89%
MobileNetV2	Yes	1.0	128	97	6.3º/87%	5.3º/90%
MobileNetV2	No	0.5	224	87	7.7º/78%	6.3º/84%
MobileNetV2	No	0.5	128	28	9.3º/68%	7.7º/74%
SqueezeNet1_1	Yes		128	87	6.3º/85%	5.6º/89%
shufflenet(1.0)	Yes		160	73	7.0º/81.62%	6.0º/85.9%