when use voxelnext to ourself dataset, the large object has poor detect effect。

[2anchao]

Hi author: ourself dataset has more large objects, like truck, The truck is over 20 meters long. below is dataset campare:

[2anchao]

    x_conv5 = self.conv5(x_conv4)
    x_conv6 = self.conv6(x_conv5)
    x_conv7 = self.conv7(x_conv6)

    x_conv5.indices[:, 1:] *= 2
    x_conv6.indices[:, 1:] *= 4
    x_conv7.indices[:, 1:] *= 8

    x_conv4 = x_conv4.replace_feature(
        torch.cat([x_conv4.features, x_conv5.features, x_conv6.features, x_conv7.features])
    )
    x_conv4.indices = torch.cat([x_conv4.indices, x_conv5.indices, 
                                 x_conv6.indices, x_conv7.indices])

    out = self.bev_out(x_conv4)

i use stride 64 feature to improve large object detect effect, and kernel_size_head=3. is it useful?any advice for me? thanks.

[yinheyanxian]

    x_conv5 = self.conv5(x_conv4)
    x_conv6 = self.conv6(x_conv5)
    x_conv7 = self.conv7(x_conv6)

    x_conv5.indices[:, 1:] *= 2
    x_conv6.indices[:, 1:] *= 4
    x_conv7.indices[:, 1:] *= 8

    x_conv4 = x_conv4.replace_feature(
        torch.cat([x_conv4.features, x_conv5.features, x_conv6.features, x_conv7.features])
    )
    x_conv4.indices = torch.cat([x_conv4.indices, x_conv5.indices, 
                                 x_conv6.indices, x_conv7.indices])

    out = self.bev_out(x_conv4)

i use stride 64 feature to improve large object detect effect, and kernel_size_head=3. is it useful?any advice for me? thanks.

does it work?