Skip to content

Latest commit

 

History

History
118 lines (91 loc) · 10.9 KB

semantic_segmentation.md

File metadata and controls

118 lines (91 loc) · 10.9 KB
Raw

Semantic segmentation

Description

Semantic segmentation is a dense prediction task that sets a goal of predicting a class label for every pixel of an input image. Generally, the objective of a segmentation algorithm is to compute a label map with the spatial resolution equal to the input.

Usual evaluation metrics are:

  • Mean Intersection over Union (mIoU): average of per-class IoU = TP/(TP + FP + FN) where TP, FP, FN - true positive, false positive, false negative predictions for pixels in evaluation set
  • Pixel Accuracy
  • Mean Average Precision
  • Combination of the two

Results

NB: evaluation of ADE20K dataset is done via calculating an average of pixel-wise accuracy and mean intersection over union.

PASCAL VOC 2012 — Validation set

Dataset description

Model Score Weights Paper Code
SDN-M2* + COCO + MS 88.80 N/A paper N/A
DeepLab_v3-Xception-65 + COCO 83.58 TensorFlow paper TensorFlow
ResNet-38 Model A1-2 + COCO 82.86 MXNet paper MXNet
ResNet-38 Model A2-2 80.84 N/A paper MXNet
SDN-M3 79.90 N/A paper N/A
DeepLab_v2 79.00 Caffe paper Caffe
ResNet-38 Model A1-2 78.76 N/A paper MXNet
ResNet-38 Model A2-1 78.14 N/A paper MXNet
DeepLab_v3-MobileNet-v2 + COCO 77.33 TensorFlow paper TensorFlow

PASCAL VOC 2012 — Testing set

Dataset description

Model Score Weights Paper Code
DeepLab_v3-Xception-65 + COCO 87.80 TensorFlow paper TensorFlow
SDN-M3 + COCO 86.60 N/A paper N/A
EncNet-ResNet-101 + COCO 85.90 N/A paper PyTorch
PSPNet-ResNet-101 + COCO 85.60 Caffe paper Caffe
ResNet-38 Model A1-2 + COCO 84.90 MXNet paper MXNet
SDN-M3 83.50 N/A paper N/A
RefineNet-ResNet-152 83.40 MatConvNet paper MatConvNet
EncNet-ResNet-101 + MS 82.90 N/A paper PyTorch
PSPNet-ResNet-101 82.60 Caffe paper Caffe
RefineNet-ResNet-101 82.40 MatConvNet paper MatConvNet
DeepLab_v3-MobileNet-v2 + COCO 80.25 TensorFlow paper TensorFlow

PASCAL Context

Dataset description

Model Score Weights Paper Code
EncNet-ResNet-101 + MS 52.60 PyTorch paper PyTorch
EncNet-ResNet-101 51.70 PyTorch paper PyTorch
ResNet-38 Model A2-2 48.10 N/A paper MXNet
RefineNet-ResNet-152 47.30 MatConvNet paper MatConvNet
RefineNet-ResNet-101 47.10 MatConvNet paper MatConvNet
DeepLab_v2 45.70 Caffe paper Caffe

ADE20K — Validation set

Dataset description

Model Score Weights Paper Code
DeepLab_v3-Xception-65 + ImageNet 64.08 TensorFlow paper TensorFlow
PSPNet-ResNet-269 + MS 63.31 N/A paper Caffe
EncNet-ResNet-101 63.17 N/A paper PyTorch
ResNet-38 Model A2-2 62.45 N/A paper MXNet
PSPNet-ResNet-269 62.34 N/A paper Caffe
ResNet-38 Model A1-2 61.56 MXNet paper MXNet
PSPNet-ResNet-50 60.86 Caffe paper Caffe
EncNet-ResNet-50 60.42 PyTorch paper PyTorch

ADE20K — Testing set

Dataset description

Model Score Weights Paper Code
ResNet-38 Model A2-2 56.41 N/A paper MXNet
EncNet-ResNet-101 + MS 55.67 N/A paper PyTorch
PSPNet-ResNet-101 55.38 N/A paper Caffe

Cityscapes — Validation set

Dataset description

Model Score Weights Paper Code
ResNet-38 Model A2-2 80.60 N/A paper MXNet
DeepLab_v3-Xception-65 + ImageNet 80.42 TensorFlow paper TensorFlow
ResNet-38 Model A1-2 78.08 MXNet paper MXNet
ResNet-38 Model A2-1 77.18 N/A paper MXNet
DeepLab_v3-MobileNet-v2 + COCO 73.57 TensorFlow paper TensorFlow

Cityscapes — Testing set

Dataset description

Model Score Weights Paper Code
DeepLab_v3-ResNet-101 + COCO + MS 82.70 N/A paper TensorFlow
PSPNet-ResNet-101 + COCO 80.20 Caffe paper Caffe
DeepLab_v3-ResNet-101 78.51 N/A paper TensorFlow
ResNet-38 Model A2-2 78.40 N/A paper MXNet
PSPNet-ResNet-101 78.40 Caffe paper Caffe
RefineNet-ResNet-101 73.60 MatConvNet paper MatConvNet