About settings for ResNet-18 backbone

[xwjabc]

Hi, recently I am training the tracker with ResNet-18 backbone. Currently, I simply reuse https://github.com/STVIR/pysot/blob/master/experiments/siamrpn_r50_l234_dwxcorr_8gpu/config.yaml and modify the

TYPE: "resnet50"
ADJUST:
    KWARGS:
        in_channels: [512, 1024, 2048]

to

TYPE: "resnet18"
ADJUST:
    KWARGS:
        in_channels: [128, 256, 512]

Since ResNet-18 uses BasicBlock which has less channels, I wonder if the channels of MultiRPN should be kept the same (i.e. RPN.KWARGS.in_channels=[256, 256]) or reduced (e.g. RPN.KWARGS.in_channels=[64, 64]). Thanks!

[ZhiyuanChen]

I'm not sure if it is quite the way to modify it to ResNet-18, but I believe as long as the in_channels of RPN matches the out_channels of ADJUST then it should be fine.

[xwjabc]

Thanks! I will try this out.

[ZhiyuanChen]

Thanks! I will try this out.

No worries, let us know the outcome