The task of 3D object detection about hr-nas HOT 9 CLOSED

Hi,

The 3D detection is based on OpenPCDet. Currently, the code is not organized into the HR-NAS repo. You can migrate our super network and loss functions to OpenPCDet for training.

We will support 3D detection in this repo as soon as possible, thank you.

Regards

from hr-nas.

First, Thanks for your reply, and I'm looking forward to your code about 3D detection.

However, I have another question. Because I saw in the paper that "FLOPs is calculated for 2D RPN network using an input size of 496×432", I would like to ask whether the 3D detection task uses SSD Detection Head when searching.

Regards

from hr-nas.

Hi,

Yes, we use a single-scale anchor-based detection head, following PointPillar. See the config for details.

Regards,
Mingyu

from hr-nas.

That is great.

    criterion = torch.nn.CrossEntropyLoss(reduction='mean').cuda()
    criterion_smooth = optim.CrossEntropyLabelSmooth(
        FLAGS.model_kwparams['num_classes'],
        FLAGS['label_smoothing'],
        reduction='mean').cuda()
    if model.task == 'segmentation':
        criterion = CrossEntropyLoss().cuda()
        criterion_smooth = CrossEntropyLoss().cuda()
    if FLAGS.dataset == 'coco':
        criterion = JointsMSELoss(use_target_weight=True).cuda()
        criterion_smooth = JointsMSELoss(use_target_weight=True).cuda()

It seems that there is not a loss for KITTI dataset in train.py, and not a yml file for KITTI in config.
Could you add them?

Thanks a lot!

from hr-nas.

Hi,

For 3D detection, we follow all settings in PointPillar (OpenPCDet reproduced version), the loss function of 3D detection can be found here.

Thanks

from hr-nas.

Hi，

Your work is very good and very helpful to my research.

I try to implement 3D detection myself, but I am not clear about many settings and it is a bit difficult.

Can you tell me how to combine your code with OpenPCDet? For example, how to prune the supernet and some search-related configurations.

Of course, it’s best if you can open source the code in the short term. If so, can you tell me how long you plan to open source the code for 3D object detection?

Thank you very much!

from hr-nas.

Hi,

Thanks for your interest. I will merge the codes as soon as time and GPU resources are available, hopefully within a month.

Regards

from hr-nas.

Feel free to reopen the issue if you have any further questions.

from hr-nas.

Hi!

Can the code be open source as scheduled？

Very much looking forward to it~~~

best regards

from hr-nas.