@@ -75,7 +75,8 @@ python main.py --image-path examples/pic1.jpg \
7575| resnet101 | ![ ] ( results/pic1-resnet101-heatmap.jpg ) | ![ ] ( results/pic1-resnet101-cam.jpg ) | ![ ] ( results/pic1-resnet101-heatmap++.jpg ) | ![ ] ( results/pic1-resnet50-cam++.jpg ) | ![ ] ( results/pic1-resnet101-gb.jpg ) | ![ ] ( results/pic1-resnet101-cam_gb.jpg ) |
7676| densenet121 | ![ ] ( results/pic1-densenet121-heatmap.jpg ) | ![ ] ( results/pic1-densenet121-cam.jpg ) | ![ ] ( results/pic1-densenet121-heatmap++.jpg ) | ![ ] ( results/pic1-densenet121-cam++.jpg ) | ![ ] ( results/pic1-densenet121-gb.jpg ) | ![ ] ( results/pic1-densenet121-cam_gb.jpg ) |
7777| inception_v3 | ![ ] ( results/pic1-inception-heatmap.jpg ) | ![ ] ( results/pic1-inception-cam.jpg ) | ![ ] ( results/pic1-inception-heatmap++.jpg ) | ![ ] ( results/pic1-inception-cam++.jpg ) | ![ ] ( results/pic1-inception-gb.jpg ) | ![ ] ( results/pic1-inception-cam_gb.jpg ) |
78- | | | | | | | |
78+ | mobilenet_v2 | ![ ] ( results/pic1-mobilenet_v2-heatmap.jpg ) | ![ ] ( results/pic1-mobilenet_v2-cam.jpg ) | ![ ] ( results/pic1-mobilenet_v2-heatmap++.jpg ) | ![ ] ( results/pic1-mobilenet_v2-cam++.jpg ) | ![ ] ( results/pic1-mobilenet_v2-gb.jpg ) | ![ ] ( results/pic1-mobilenet_v2-cam_gb.jpg ) |
79+ | shufflenet_v2 | ![ ] ( results/pic1-shufflenet_v2-heatmap.jpg ) | ![ ] ( results/pic1-shufflenet_v2-cam.jpg ) | ![ ] ( results/pic1-shufflenet_v2-heatmap++.jpg ) | ![ ] ( results/pic1-shufflenet_v2-cam++.jpg ) | ![ ] ( results/pic1-shufflenet_v2-gb.jpg ) | ![ ] ( results/pic1-shufflenet_v2-cam_gb.jpg ) |
7980
8081### 多个对象
8182
@@ -95,11 +96,14 @@ python main.py --image-path examples/pic1.jpg \
9596| resnet101 | ![ ] ( results/multiple_dogs-resnet101-heatmap.jpg ) | ![ ] ( results/multiple_dogs-resnet101-cam.jpg ) | ![ ] ( results/multiple_dogs-resnet101-heatmap++.jpg ) | ![ ] ( results/multiple_dogs-resnet50-cam++.jpg ) | ![ ] ( results/multiple_dogs-resnet101-gb.jpg ) | ![ ] ( results/multiple_dogs-resnet101-cam_gb.jpg ) |
9697| densenet121 | ![ ] ( results/multiple_dogs-densenet121-heatmap.jpg ) | ![ ] ( results/multiple_dogs-densenet121-cam.jpg ) | ![ ] ( results/multiple_dogs-densenet121-heatmap++.jpg ) | ![ ] ( results/multiple_dogs-densenet121-cam++.jpg ) | ![ ] ( results/multiple_dogs-densenet121-gb.jpg ) | ![ ] ( results/multiple_dogs-densenet121-cam_gb.jpg ) |
9798| inception_v3 | ![ ] ( results/multiple_dogs-inception-heatmap.jpg ) | ![ ] ( results/multiple_dogs-inception-cam.jpg ) | ![ ] ( results/multiple_dogs-inception-heatmap++.jpg ) | ![ ] ( results/multiple_dogs-inception-cam++.jpg ) | ![ ] ( results/multiple_dogs-inception-gb.jpg ) | ![ ] ( results/multiple_dogs-inception-cam_gb.jpg ) |
98- | | | | | | | |
99+ | mobilenet_v2 | ![ ] ( results/multiple_dogs-mobilenet_v2-heatmap.jpg ) | ![ ] ( results/multiple_dogs-mobilenet_v2-cam.jpg ) | ![ ] ( results/multiple_dogs-mobilenet_v2-heatmap++.jpg ) | ![ ] ( results/multiple_dogs-mobilenet_v2-cam++.jpg ) | ![ ] ( results/multiple_dogs-mobilenet_v2-gb.jpg ) | ![ ] ( results/multiple_dogs-mobilenet_v2-cam_gb.jpg ) |
100+ | shufflenet_v2 | ![ ] ( results/multiple_dogs-shufflenet_v2-heatmap.jpg ) | ![ ] ( results/multiple_dogs-shufflenet_v2-cam.jpg ) | ![ ] ( results/multiple_dogs-shufflenet_v2-heatmap++.jpg ) | ![ ] ( results/multiple_dogs-shufflenet_v2-cam++.jpg ) | ![ ] ( results/multiple_dogs-shufflenet_v2-gb.jpg ) | ![ ] ( results/multiple_dogs-shufflenet_v2-cam_gb.jpg ) |
99101
100102
101103
102104## 总结
103105
104106- vgg模型的Grad-CAM并没有覆盖整个对象,相对来说resnet和denset覆盖更全,特别是densenet;从侧面说明就模型的泛化和鲁棒性而言densenet>resnet>vgg
105- - Grad-CAM++相对于Grad-CAM也是覆盖对象更全面，特别是对于同一个类别有多个实例的情况下,Grad-CAM可能只覆盖部分对象，Grad-CAM++基本覆盖所有对象;但是这仅仅对于vgg而言,想densenet直接使用Grad-CAM也基本能够覆盖所有对象
107+ - Grad-CAM++相对于Grad-CAM也是覆盖对象更全面，特别是对于同一个类别有多个实例的情况下,Grad-CAM可能只覆盖部分对象，Grad-CAM++基本覆盖所有对象;但是这仅仅对于vgg而言,想densenet直接使用Grad-CAM也基本能够覆盖所有对象
108+ - MobileNet V2的Grad-CAM覆盖也很全面
109+ - Inception V3和MobileNet V2的Guided backpropagation图轮廓很模糊，但是ShuffleNet V2的轮廓则比较清晰
0 commit comments