replace vgg with resnet

[lyxbyr]

Hi Thanks for your great work, I am now trying to replace the backbone of second stage using Resnet since Resnet usually has a better performance when doing a same work. However, after I replace the vgg by resnet, the result is terrible. I wonder if you have do the same thing and cloud you please tell me whether this idea is useful. Thank you!

Best regards

[cnexah]

Hi, I also tried Resnet18 and Resnet50, but got a similar performance to VGG.

[lyxbyr]

@Cknex Hi ， similar performance how to eval it ? computer dimension error and orientation error ？the final three branch fc layer node need to change,eg resnet18 or resnet50 the three branch fc you set how many fc node?

[lyxbyr]

my result: Angle error Dimension error Vgg19 13.683045 0.283335 resnet34 27.887893 0.849132 resnet50 21.706590 0.821558

[cnexah]

Sorry, I made a mistake. My performance of Resnet50 is also bad.

I use the mean absolute error, and only test for 'Car' Angle error, Dimension error VGG19 0.108 [0.086 0.067 0.194] Resnet50 0.534 [0.100 0.076 0.320]

As for the FC layers, I just simply adapt the first fc layer to be compatible with the output of Resnet50

[lyxbyr]

Hi,for resnet50,fc like this result may be better : self.orientation = nn.Sequential( nn.Linear(2048 7 7, 1024), nn.ReLU(True), nn.Dropout(), nn.Linear(1024, 1024), nn.ReLU(True), nn.Dropout(), nn.Linear(1024, bins 2) # to get sin and cos ) self.confidence = nn.Sequential( nn.Linear(2048 7 * 7, 1024), nn.ReLU(True), nn.Dropout(), nn.Linear(1024, 1024), nn.ReLU(True), nn.Dropout(), nn.Linear(1024, bins),

nn.Softmax()

        # nn.Sigmoid()
    )
    self.dimension = nn.Sequential(
        nn.Linear(2048 * 7 * 7, 512),
        nn.ReLU(True),
        nn.Dropout(),
        nn.Linear(512, 512),
        nn.ReLU(True),
        nn.Dropout(),
        nn.Linear(512, 3)
    )

[cnexah]

Thanks very much!

[CaoBo1996]

@Cknex hey，i found your result Angle error, Dimension error VGG19 0.108 [0.086 0.067 0.194] Resnet50 0.534 [0.100 0.076 0.320]

Angle error was very good,but in my test,my result is bad. i split kitti train set into split_train and split_val i train in split_train，and test in split_val but my result is bad i also use mean absolute error,and test in Car this is my result Car angle error：0.4860737579686576 so can you help me why?

[cnexah]

@CaoBo1996 Sorry that I have lost my code. In memory I didn't change much. Maybe just more units in middle layers.