3DMatch model on KITTI

[fabiopoiesi]

Hi,

I am trying to use the 3DMatch model on KITTI, but so far I only get 0% succ rate on KITTI. In addition to setting first_subsampling_dl = 0.30, do you have any other recommendation to improve performance?

Thanks

[XuyangBai]

Hi, thanks for your interest. It is similar to the experiments that evaluate the 3DMatch model on ETH, you should also change the scale of kernel point as follow https://github.com/XuyangBai/D3Feat/blob/3577482efbc5154affcd734e16b0d10a73560e37/demo_registration.py#L139-L148

You may find some useful discussion https://github.com/XuyangBai/D3Feat/issues/10 and https://github.com/XuyangBai/D3Feat/issues/14. In general D3Feat (as well as other fully-convolution descriptors) may not have good generalization ability as patch-based ones, but it should not be 0%.

Best, Xuyang

[fabiopoiesi]

Thank you for your prompt response.

I uncommentend these lines of code here and relaunched the script but I still have 0% succ rate.

Here below the steps I did.

Before this line I added chosen_log = 'results/Log_contraloss/' and modified the following lines as: line -> dataset = 'KITTI' line -> first_subsampling_dl = 0.30

If I use the same script (test_kitti.py) with the KITTI checkpoint it works without problems (of course without the modifications above).

[XuyangBai]

Hi, did you enlarge the scale by 0.10/0.03 or 0.30/0.03?

https://github.com/XuyangBai/D3Feat/blob/476df5362bb398a0104266f4d1598cc54de21712/utils/tester.py#L164-L170

I think this hyperparameter and the first_subsampling_dlmay have a large effect on the performance, maybe play with these two hyperparameters can give you better results. Otherwise I also have no idea on how to make the generalization work, I have not tried such experiments.

[fabiopoiesi]

Using 0.30/0.03 it works! Thanks!!

[XuyangBai]

Cool:) Glad to hear that.

[fabiopoiesi]

How can I make the same modification to the pytorch version?

[XuyangBai]

You can scale up the kernel points in a similar way:

    model = KPFCNN(config)
    weights = torch.load(f'snapshot/{args.chosen_snapshot}/models/model_best_acc.pth', map_location='cpu')['state_dict']
    for key, value in weights.items():
        if 'kernel_points' in key:
            weights[key] = value * config.first_subsampling_dl / 0.03
    model.load_state_dict(weights)

[fabiopoiesi]

It also works in pytorch. Thanks for your support.