The scale of response map. about siamdw HOT 4 CLOSED

Thanks for your great work! It helps me a lot.

The code of convolution layer of Siamfc model is:
def forward(self, z_f, x_f): if not self.training: return 0.1 * F.conv2d(x_f, z_f) else: return 0.1 * self._conv2d_group(x_f, z_f)

Why change the scale of the results of cross-correlation? How to determine parameter 0.1?

You can use the weight learned from network(a 1*1 conv layer) or simply give a scale factor. There is no obvious difference on performance. Any other question?

from siamdw.

Thanks for your great work! It helps me a lot.
The code of convolution layer of Siamfc model is:
def forward(self, z_f, x_f): if not self.training: return 0.1 * F.conv2d(x_f, z_f) else: return 0.1 * self._conv2d_group(x_f, z_f)
Why change the scale of the results of cross-correlation? How to determine parameter 0.1?

You can use the weight learned from network(a 1*1 conv layer) or simply give a scale factor. There is no obvious difference on performance. Any other question?

Sorry, I still don't understand. If there is no difference of them, why do that? How about use the results of cross-correlation directly like this:
def forward(self, z_f, x_f): if not self.training: return F.conv2d(x_f, z_f) else: return * self._conv2d_group(x_f, z_f) ?

from siamdw.

Thanks for your great work! It helps me a lot.
The code of convolution layer of Siamfc model is:
def forward(self, z_f, x_f): if not self.training: return 0.1 * F.conv2d(x_f, z_f) else: return 0.1 * self._conv2d_group(x_f, z_f)
Why change the scale of the results of cross-correlation? How to determine parameter 0.1?

You can use the weight learned from network(a 1*1 conv layer) or simply give a scale factor. There is no obvious difference on performance. Any other question?

Sorry, I still don't understand. If there is no difference of them, why do that? How about use the results of cross-correlation directly like this:
def forward(self, z_f, x_f): if not self.training: return F.conv2d(x_f, z_f) else: return * self._conv2d_group(x_f, z_f) ?

We usually use a 1*1 conv layer to adjust the scale of correlation output. The output of correlation is usually very large at the beginning of training phase, a small scale factor is beneficial to network convergence.

from siamdw.

Thanks for your great work! It helps me a lot.
The code of convolution layer of Siamfc model is:
def forward(self, z_f, x_f): if not self.training: return 0.1 * F.conv2d(x_f, z_f) else: return 0.1 * self._conv2d_group(x_f, z_f)
Why change the scale of the results of cross-correlation? How to determine parameter 0.1?

You can use the weight learned from network(a 1*1 conv layer) or simply give a scale factor. There is no obvious difference on performance. Any other question?

Sorry, I still don't understand. If there is no difference of them, why do that? How about use the results of cross-correlation directly like this:
def forward(self, z_f, x_f): if not self.training: return F.conv2d(x_f, z_f) else: return * self._conv2d_group(x_f, z_f) ?

We usually use a 1*1 conv layer to adjust the scale of correlation output. The output of correlation is usually very large at the beginning of training phase, a small scale factor is beneficial to network convergence.

Thanks.

from siamdw.