deepvac_human_accept600 SOTA

[gemfield]

IoU 金榜

• test9: 91.5%
• test5: 90.9%
• test6: 90.5%
• test7: 90.5%
• test3: 87.2%
• test2: 86.9%
• test4: 86.3%
• test1: 85.7%
• test8: 85.6%

[1icas]

test1

IoU

• label: 85.7%
• label_no_mix: ??

特别超参

• 训练集：HumanSeg-Mapillary
• config.input_w = 224
• config.input_h = 384
• config.core.optimizer = torch.optim.SGD(config.core.net.parameters(), 0.01, momentum=0.9)
• config.core.scheduler = optim.lr_scheduler.MultiStepLR(config.core.optimizer, milestones=[4, 8, 12, 15], gamma=0.1)

基本超参

config.core.cls_num = 2
config.core.batch_size = 64
config.core.mean = config.data['mean']
config.core.std = config.data['std']
config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
config.core.criterion = torch.nn.CrossEntropyLoss(weight)
AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)

[1icas]

test2

IoU

• label: 86.9%
• label_no_mix: ??

特别超参

• 训练集：HumanSeg-Mapillary + 1倍HumanSeg-Mapillary 背景替换
• config.input_w = 224
• config.input_h = 384
• config.core.optimizer = torch.optim.SGD(config.core.net.parameters(), 0.01, momentum=0.9)
• config.core.scheduler = optim.lr_scheduler.MultiStepLR(config.core.optimizer, milestones=[4, 8, 12, 15], gamma=0.1)

基本超参

  config.core.cls_num = 2
  config.core.batch_size = 64
  config.core.mean = config.data['mean']
  config.core.std = config.data['std']
  config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
  config.core.criterion = torch.nn.CrossEntropyLoss(weight)
  AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)

[1icas]

test3

IoU

• label: 87.2%
• label_no_mix: ??

特别超参

• 训练集：HumanSeg-Mapillary + 4倍HumanSeg-Mapillary 背景替换
• config.input_w = 224
• config.input_h = 384
• config.core.optimizer = torch.optim.SGD(config.core.net.parameters(), 0.01, momentum=0.9)
• config.core.scheduler = optim.lr_scheduler.MultiStepLR(config.core.optimizer, milestones=[4, 8, 12, 15], gamma=0.1)

基本超参

  config.core.cls_num = 2
  config.core.batch_size = 64
  config.core.mean = config.data['mean']
  config.core.std = config.data['std']
  config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
  config.core.criterion = torch.nn.CrossEntropyLoss(weight)
  AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)

[1icas]

test4

IoU

• label: 86.3%
• label_no_mix: ??

特别超参

• 训练集：HumanSeg-Mapillary + 4倍HumanSeg-Mapillary 背景替换 + 2w张纯背景
• config.input_w = 224
• config.input_h = 384
• config.core.optimizer = torch.optim.SGD(config.core.net.parameters(), 0.01, momentum=0.9)
• config.core.scheduler = optim.lr_scheduler.MultiStepLR(config.core.optimizer, milestones=[4, 8, 12, 15], gamma=0.1)

基本超参

  config.core.cls_num = 2
  config.core.batch_size = 64
  config.core.mean = config.data['mean']
  config.core.std = config.data['std']
  config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
  config.core.criterion = torch.nn.CrossEntropyLoss(weight)
  AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)

[1icas]

test5

IoU

• label: 90.9%
• label_no_mix:90.8%

特别超参

• 训练集：HumanSeg-Mapillary + CoCo + 2.5w张纯背景
• config.core.optimizer = torch.optim.SGD(config.core.net.parameters(), 0.01, momentum=0.9, weight_decay=1e-5)
• config.core.scheduler = optim.lr_scheduler.MultiStepLR(config.core.optimizer, milestones=[6, 12, 15, 18], gamma=0.1)

模型修改

• 去掉了网络中所有的nn.Dropout2d

基本超参

config.input_w = 384
config.input_h = 384
config.core.cls_num = 2
config.core.batch_size = 16
config.core.mean = config.data['mean']
config.core.std = config.data['std']
config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
config.core.criterion = torch.nn.CrossEntropyLoss(weight)
AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)

[MHGL]

test6

IoU

• epoch: 93
• labels: 90.5%
• labels_no_mix: 90.1%

特别超参

• 训练集： deepvac-chengdu(1526 no mix) + deepvac-waibao(1589 no mix)
• config.input_w = 384
• config.input_h = 384
• config.core.criterion = torch.nn.CrossEntropyLoss(weight)

基本超参

config.core.epoch_num = 100
config.core.cls_num = 2
config.core.batch_size = 16
config.core.mean = config.data['mean']
config.core.std = config.data['std']
config.core.optimizer = torch.optim.Adam(config.core.net.parameters(), 3e-4, (0.9, 0.999), eps=1e-08, weight_decay=5e-4)
lambda_lr = lambda epoch: round ((1 - epoch/config.core.epoch_num) ** 0.9, 8)
config.core.scheduler = optim.lr_scheduler.LambdaLR(config.core.optimizer, lr_lambda=lambda_lr)
config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)

[MHGL]

test7

IoU

• epoch: 95
• labels: 90.5%
• labels_no_mix: 90.3%

特别超参

• 训练集： deepvac-chengdu(1526 no mix) + deepvac-waibao(1589 no mix)
• config.input_w = 384
• config.input_h = 384
• config.core.criterion = bce_loss + dice_loss

基本超参

config.core.epoch_num = 100
config.core.cls_num = 2
config.core.batch_size = 16
config.core.mean = config.data['mean']
config.core.std = config.data['std']
config.core.optimizer = torch.optim.Adam(config.core.net.parameters(), 3e-4, (0.9, 0.999), eps=1e-08, weight_decay=5e-4)
lambda_lr = lambda epoch: round ((1 - epoch/config.core.epoch_num) ** 0.9, 8)
config.core.scheduler = optim.lr_scheduler.LambdaLR(config.core.optimizer, lr_lambda=lambda_lr)
config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)

[MHGL]

test8

IoU

• 前sota模型
• labels: 85.6%
• labels_no_mix: 85.5%

特别超参

• 训练集： deepvac-chengdu(1526 no mix) + supervisely(5711)
• config.input_w = 384
• config.input_h = 384
• config.core.criterion = bce_loss + dice_loss

基本超参

config.core.epoch_num = 100
config.core.cls_num = 2
config.core.batch_size = 16
config.core.mean = config.data['mean']
config.core.std = config.data['std']
config.core.optimizer = torch.optim.Adam(config.core.net.parameters(), 3e-4, (0.9, 0.999), eps=1e-08, weight_decay=5e-4)
lambda_lr = lambda epoch: round ((1 - epoch/config.core.epoch_num) ** 0.9, 8)
config.core.scheduler = optim.lr_scheduler.LambdaLR(config.core.optimizer, lr_lambda=lambda_lr)
config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"

[MHGL]

test9

IoU

• epoch: 71
• labels: 91.5%
• labels_no_mix: 91.4%

特别超参

• 训练集： deepvac-chengdu(1646 no mix) + deepvac-waibao(4399 no mix) + onlyBG(1000)
• config.input_w = 384
• config.input_h = 384
• config.core.criterion = CrossEntropyLoss

基本超参

config.core.epoch_num = 100
config.core.cls_num = 2
config.core.batch_size = 16
config.core.mean = config.data['mean']
config.core.std = config.data['std']
config.core.optimizer = torch.optim.Adam(config.core.net.parameters(), 3e-4, (0.9, 0.999), eps=1e-08, weight_decay=5e-4)
lambda_lr = lambda epoch: round ((1 - epoch/config.core.epoch_num) ** 0.9, 8)
config.core.scheduler = optim.lr_scheduler.LambdaLR(config.core.optimizer, lr_lambda=lambda_lr)
config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)