FCN experiment

[yzbx]

official implement

backbone network is vgg16 backbone pretrained = true backbone freeze = false batch_size=20 input size=500x500 optimizer = sgd(1e-4) + momentum(0.9) weight decay=5^-4 or 2^-4 double learning rate for bias (not recommend) zero-initialize the class scoring layer (not recommend) train 3 days in single gpu for fcn32(miou=59.4) finetune fcn32 for 1 day, got fcn16(miou=62.4) finetune fcn16 for 1 day, got fcn8(miou=62.7) upsample type = bilinear + deconv(init by bilinear)

official finds

class balancing unnecessary augmentation yielded no noticeable improvement training sensitive to learning rate alone

fcn series

upper bound miou and downsample factor

[yzbx]

my implement

python test/pspnet_test.py --backbone_name=vgg16_gn --use_momentum=True --batch_size=4 --upsample_layer=3 --midnet_scale=10 --use_lr_mult=False --net_name=fcn --note=fcn8_vgg16

python test/pspnet_test.py --backbone_name=vgg16_gn --use_momentum=True --batch_size=4 --upsample_layer=4 --midnet_scale=10 --use_lr_mult=False --net_name=fcn --note=fcn16_vgg16

python test/pspnet_test.py --backbone_name=vgg16_gn --use_momentum=True --batch_size=4 --upsample_layer=5 --midnet_scale=10 --use_lr_mult=False --net_name=fcn --note=fcn32_vgg16

network	mean iou
fcn8_vgg16	0.482

[dishen12]

vgg19 version (remove maxpool after layer 3)

network	miou	note
fcn8_vgg19	0.498	fcn8_vgg19
fcn16_vgg19	0.577	fcn8_vgg19
fcn32_vgg19	0.588	fcn8_vgg19

use standard fcn suffix layer

python test/pspnet_test.py --backbone_name=vgg19_gn --use_none_layer=True --batch_size=4 --upsample_layer=5 --upsample_type=fcn --midnet_scale=10 --use_lr_mult=False --net_name=fcn --note=fcn32_vgg16_fcn

vgg19 version (keep maxpool after layer 3)

python test/pspnet_test.py --backbone_name=vgg19_gn --use_none_layer=True --batch_size=4 --upsample_layer=3,4,5 --midnet_scale=10 --use_lr_mult=False --net_name=fcn --note=fcn_vgg19

use standard fcn suffix layer

python test/pspnet_test.py --backbone_name=vgg19_gn --use_none_layer=True --batch_size=4 --upsample_layer=3,4,5 --upsample_type=fcn --midnet_scale=10 --use_lr_mult=False --net_name=fcn --note=fcn_vgg19_fcn

[yzbx]

	train/iou	val/iou	note	backbone_name	upsample_layer	midnet_scale	n_epoch
3	0.837138	0.408618	vgg19_bn_psp	vgg19_bn	4	10	100
14	0.55358	0.473572	vgg16_fz_psp	vgg16	5	10	100
2	0.602388	0.481789	fcn8_vgg16	vgg16_gn	3	10	100
15	0.578943	0.487695	cls_wise_vgg16	vgg16	5	15	100
12	0.684266	0.523931	vgg11_psp	vgg11	4	10	100
10	0.664645	0.539331	vgg19_gn	vgg19_gn	3	10	100
7	0.666785	0.539916	vgg19_gn_nopre	vgg19_gn	3	15	100
1	0.711822	0.560617	fcn16_vgg16	vgg16_gn	4	10	100
11	0.808483	0.578575	lr4_lr20_vgg16	vgg16	5	15	100
9	0.70724	0.580211	vgg16_gn	vgg16_gn	4	10	100
16	0.797704	0.585897	lr4_lr20_vgg19	vgg19	5	15	100
4	0.71189	0.588305	vgg16_gn32	vgg16_gn	4	10	100
6	0.72956	0.589235	vgg16_bn	vgg16_bn	4	10	100
0	0.722444	0.592189	fcn32_vgg16	vgg16_gn	5	10	100
5	0.668338	0.593148	vgg19_freeze3_5	vgg19_bn	5	10	100
8	0.704294	0.599245	vgg19_gn_nopre	vgg19_gn	4	15	100
13	0.701463	0.622954	vgg16_duc	vgg16_bn	4	10	100

	train/iou	val/iou	note	backbone_name	upsample_layer	midnet_scale	n_epoch
2	0.602388	0.481789	fcn8_vgg16	vgg16_gn	3	10	100
1	0.711822	0.560617	fcn16_vgg16	vgg16_gn	4	10	100
0	0.722444	0.592189	fcn32_vgg16	vgg16_gn	5	10	100