DeepCSF

Contrast Sensitivity Function (CSF) in deep networks

The contrast sensitivity function (CSF) is a fundamental signature of the visual system that has been measured extensively in several species. It is defined by the visibility threshold for sinusoidal gratings at all spatial frequencies. Here, we investigated the CSF in deep neural networks using the same 2AFC contrast detection paradigm as in human psychophysics.

Source Code Article

Results

  1. ImageNet
  2. Without Pretrained Weights
  3. Taskonomy
  4. MS COCO
  5. CLIP

ImageNet

ResNet

ResNet18

resnet18_i1

ResNet34

resnet34_i1

ResNet50

resnet50_i0
resnet50_i1
resnet50_i2
resnet50_i3
resnet50_i4
resnet50_i5
resnet50_i6
resnet50_i7
resnet50_i8
resnet50_i9

ResNet101

resnet101_i1

ResNet152

resnet152_i1

ResNeXt

resnext50_32x4d_i1
resnext101_32x8d_i1

Wide-ResNet

wide_resnet50_2_i1
wide_resnet101_2_i1

ViT

ViT-B32

vit_b_32_i0
vit_b_32_i1
vit_b_32_i2
vit_b_32_i3
vit_b_32_i4
vit_b_32_i5
vit_b_32_i6
vit_b_32_i7
vit_b_32_i8
vit_b_32_i9

ViT-L32

vit_l_32_i0
vit_l_32_i1
vit_l_32_i2
vit_l_32_i3
vit_l_32_i4
vit_l_32_i5
vit_l_32_i6
vit_l_32_i7
vit_l_32_i8
vit_l_32_i9

ConvNeXt

convnext_tiny_i0
convnext_tiny_i1
convnext_tiny_i2
convnext_tiny_i3
convnext_tiny_i4
convnext_tiny_i5
convnext_tiny_i6
convnext_tiny_i7
convnext_tiny_i8
convnext_tiny_i9

RegNet

regnet_x_1_6gf_i0
regnet_x_1_6gf_i1
regnet_x_1_6gf_i2
regnet_x_1_6gf_i3
regnet_x_1_6gf_i4
regnet_x_1_6gf_i5
regnet_x_1_6gf_i6
regnet_x_1_6gf_i7
regnet_x_1_6gf_i8
regnet_x_1_6gf_i9

VGG

VGG16-BN

vgg16_bn_i0
vgg16_bn_i1
vgg16_bn_i2
vgg16_bn_i3
vgg16_bn_i4
vgg16_bn_i5
vgg16_bn_i6
vgg16_bn_i7
vgg16_bn_i8
vgg16_bn_i9

VGG16

vgg16_i0
vgg16_i1
vgg16_i2
vgg16_i3
vgg16_i4
vgg16_i5
vgg16_i6
vgg16_i7
vgg16_i8
vgg16_i9

Classification

alexnet_i1
convnext_base_i1
convnext_large_i1
convnext_small_i1
densenet121_i1
densenet161_i1
densenet169_i1
densenet201_i1
efficientnet_b0_i1
efficientnet_b1_i1
efficientnet_b2_i1
efficientnet_b3_i1
efficientnet_b4_i1
efficientnet_b5_i1
efficientnet_b6_i1
efficientnet_b7_i1
googlenet_i1
inception_v3_i1
mnasnet0_5_i1
mnasnet1_0_i1
mobilenet_v2_i1
mobilenet_v3_large_i1
mobilenet_v3_small_i1
regnet_x_16gf_i1
regnet_x_32gf_i1
regnet_x_3_2gf_i1
regnet_x_400mf_i1
regnet_x_800mf_i1
regnet_x_8gf_i1
regnet_y_16gf_i1
regnet_y_1_6gf_i1
regnet_y_32gf_i1
regnet_y_3_2gf_i1
regnet_y_400mf_i1
regnet_y_800mf_i1
regnet_y_8gf_i1
shufflenet_v2_x0_5_i1
shufflenet_v2_x1_0_i1
squeezenet1_0_i1
squeezenet1_1_i1
vgg11_bn_i1
vgg11_i1
vgg13_bn_i1
vgg13_i1
vgg19_bn_i1
vgg19_i1
vit_b_16_i1
vit_l_16_i1

Linear-SVM

resnet50_i0
resnet50_i1
resnet50_i2
resnet50_i3
resnet50_i4
resnet50_i5

Without-pretrained-weights

resnet50_random_weights_i0
resnet50_random_weights_i1
resnet50_random_weights_i2
resnet50_random_weights_i3
resnet50_random_weights_i4
resnet50_random_weights_i5
resnet50_random_weights_i6
resnet50_random_weights_i7
resnet50_random_weights_i8
resnet50_random_weights_i9

Taskonomy

autoencoding
class_object
class_scene
curvature
denoising
depth_euclidean
depth_zbuffer
edge_occlusion
edge_texture
egomotion
fixated_pose
inpainting
jigsaw
keypoints2d
keypoints3d
nonfixated_pose
normal
point_matching
reshading
room_layout
segment_semantic
segment_unsup25d
segment_unsup2d
vanishing_point

MS COCO

deeplabv3_resnet101
deeplabv3_resnet50
fcn_resnet101
fcn_resnet50

CLIP

B16_i1
L14_i1
RN101_i1

CLIP-B32

B32_i0
B32_i1
B32_i2
B32_i3
B32_i4
B32_i5
B32_i6
B32_i7
B32_i8
B32_i9

CLIP-RN50

RN50_i0
RN50_i1
RN50_i2
RN50_i3
RN50_i4
RN50_i5
RN50_i6
RN50_i7
RN50_i8
RN50_i9