yinjunbo/proposalcontrast

This repository contains the PyTorch implementation of the ECCV'2022 paper, ProposalContrast: Unsupervised Pre-training for LiDAR-based 3D Object Detection.

proposalcontrast's Issues

Can you provide the pretrained model checkpoints?

As the title said.

where is the part of cross-attention mechanism ?

Hello, is the code contains the mentioned cross-attention in the paper?

How many data are used for self-supervised pretraining?

Hi Junbo,
Thanks for your great work and open-sourced code. I have a problem of the amount of training data used in self-supervised pretraining. You mentioned in the paper that the whole training set of Waymo is used for pretraining. However, I notice the config file for SSL that the "load_interval" is set to 4, which means that only 25% training data are used during pretraining.

ProposalContrast/configs/waymo/voxelnet/ssl_pretrain_voxelnet.py

Line 113 in 3b02290

load_interval=4

How did you sample the datasets for data-efficient 3d detection

Hi, thank you for your excellent work.

In your paper about data-efficient 3D object detection on the Waymo dataset, you said "During fine-tuning, various fractions of training data are uniformly sampled". I want to ask whether you sample the data frame by frame or scene by scene.
Frame by Frame means all scenes(399 scenes in your paper) are included but every (for example for 10%) 10 frames you sample 1.
Scene by scene means you sample some scenes(40 scenes for 10%) and all the frames in those scenes are used for fine-tuning.

Also, how do you split the training samples of KITTI into three groups?

Thank you for your answer.

Question about point drop out and key point selection

Hello, I'm a student who has a great interest in your work.

There is a simple question about the key point selection.

When you drop out some points and retain others, there are overlapped points between point clouds X1 and X2.

Do the overlapped points include key points that have been selected by FPS?

Thank you.

Unable to download ground plane data

Hello,
Is there a way to download the ground plane information without a Baidu account?

Is it necessary to remove road points when pre-training?

Hi, thank you for your excellent work.
I am trying to reimplement your code and my codebase does not support removing points. I am wondering if it is necessary to remove the road points when pre-training. Will not removing hurt the performance?
Thank you.

[Bug?] in forward_ssl

Hi Junbo,

In your forward_ssl code, you squeeze the fps sampling sequence, but this will raise errors when the batch size is 1 because you index the fps_choice with the batch index afterward.

    fps_choice = pointnet2_batch_utils.furthest_point_sample(
            points_overlap_org.contiguous(), self.npoints
        ).long().squeeze()

        # compute correspondence
        for kk in range(batch_size):
            key_idx = example['overlap'][kk][fps_choice[kk]].type(torch.float32)
            example['correspondence'][kk] = torch.min(torch.abs(torch.sub(key_idx.unsqueeze(dim=-1),example['correspondence'][kk].unsqueeze(dim=0).type(torch.float32))),dim=-1)[1]

Could you release the code for getting the ground points?

Hi, Junbo. Thanks a lot for sharing your great work. Could you please release the code for getting the ground points? I want to use it on other datasets, e.g., Nuscenes. Or could you give more detail about it?

How do you re-implement PointContrast and DepthContrast?

Hi, thank you for your excellent work.

I wonder how you re-implement PointContrast and DephtContrast in your codebase? Can you give me some hints?

If you can provide the code (or some code snippets), I will appreciate it very much.

Solved

2022-09-15 00:54:55,076 - INFO - Reducer buckets have been rebuilt in this iteration.

how to fix this problem

The shape mismatch for lidar_ground segmentation array and points

Hi Junbo,

Thanks for your great work!
When I run the pertaining for waymo, I found the shape mismatch for lidar_ground segmentation array and points. The largest value in lidar_ground is even larger than the length of the points... I wonder how can this happen?
Thanks,
Mu

baseline results

Hi,
Thanks for the great work!
I have one question about the baseline results of Table 2 in the paper. The paper claims that the baseline results on Waymo Dataset are obtained following OpenPCDet. However, I find the results reported in the paper are much lower than that in OpenPCDet. For example, the result of CenterPoint (VoxelNet) is 61.81/61.30, 63.62/57.79, and 64.96/63.77 in the paper. OpenPCDet reports the result of 63.16/62.65, 64.27/58.23, and 66.11/64.87 for CenterPoint (VoxelNet).

code release?

Hi, authors. Thanks for your great work.

Do you have exact date for releasing the code?