rsy6318 / corri2p

[TCSVT] CorrI2P: Deep Image-to-Point Cloud Registration via Dense CorrespondenceThe code of CorrI2P

corri2p's Issues

Data of NuScenes.

Hi,
Do you have the plan on releasing the dataset and dataloader for NuScenes experiments. Since I would like to make a more complete comparison with your great work.
Thanks for releasing your code, looking forward to your reply.

About success rate of CorrI2P

Hi Siyu, thanks for your nice work.
I noticed that the paper seems to evaluate samples with rte and rre less than a certain threshold.
Is it convenient for you to provide the quantitative results of the full sample and the success rate of RansacPnP?
Looking forward to your reply!

How to picture the correspondence_ours.png

thank you share your code, i want to know how to draw the result with line

Can not download the processed NuScenes datasets

Hi Siyu, thanks for your nice work.
I noticed that the link to download the processed NuScenes datasets seems to go wrong for some reason.
Is it convenient for you to provide a new link to get the processed NuScenes datasets?
Looking forward to your reply! Thanks!

how to install the 'index_max'?

Thanks for sharing your works， when i try to install the index_max follow SO-Net instructions，but still got the following error：

I use the pycharm compiler, the environment is pytorch3.9, cuda11.6, windows

The scripts of the processing of NuScenes

Dear authors,
As mentioned in issue #3, you have updated the NuScenes dataset processed. Thanks for your effort for this. However, the scripts about NuScenes you updated (in the directory nuScenes_script) don't include the processing code from the raw data of NuScenes to the data form used in your dataloader script nuscenes_pc_img_dataloader.py . Specifically, the data form with the subfix .npy, which is included in the processed dataset you updated. So could you please update the processing scripts for the NuScenes dataset? This would help me a lot and make me understand more detail about the description We used the official SDK to get the image-point cloud pairs, where the point cloud was accumulated from the nearby frames, and the image from the current data frame in your paper.

Thanks for your nice work and releasing of your experiment code.

ModuleNotFoundError: No module named 'index_max'

I dont know how to use index_max. I tried to use your link about DEEPI2P, but it doesnt work. What I can do?

about index_max

I try to run the code , but i cant install index_max, no mattter use pip or conda, how can I solve it

Kitti dataset failing to download

Hello,

I can't download the kitti dataset.
Can you please provide a script to adapt the original dataset to your required format?

How can I get 'K', 'K_P2', 'K_P3'?

Hi, thanks for your open-sourced work.
I have converted the official KITTI images and LiDAR data to .npy form by you scripts, but I can't run the code successfully because I lack the pose related files. I only find ten .txt files in KITTI offical website and no .npz files. Can you tell me how can I get 'K', 'K_P2', 'K_P3'?
BTW, I can't download and unzip your uploaded data (I have tried many times). @rsy6318

I noticed that you have a description similar to outline_index in your code （in kitti_pc_img_dataloader.py)

Does outline refer to the outer contour
or
points in the point cloud that exceed the image range?

Descriptor loss during training.

Hi,
Thanks again for sharing your great job! I processed the Kitti data myself and reproduced CorrI2P. I find the descriptor loss is hard to drop and is around 3.8 and I am not sure if it's normal. so could you share your pretrained model and training log in Kitti dataset for my reference on the correctness of my reproduction.

Code releasing.

Hi,
Thanks for your great job! And when will you release the code?

About data generation of nuScenes.

Hi,
I am currently visualizing the images of nuScenes, but I found that when I generate a video with a set of image frames(e.g. 0-1000), the video route is not always forward, it is very concussive. However, the route on the KITTI dataset is quite normal (when I only adopt the images from left(or right) camera). I think it's the different way for generating KITTI and nuScenes dataset (is there many cameras on the nuScenes dataset and you put their captured images together?)

Looking forward to your reply, thanks for your help!

Pretrained model file

Hello, can you provide the pre-trained model file?

ImportError: cannot import name 'DenseI2P' from 'network'

There is an error when run evel_all.py: ImportError: cannot import name 'DenseI2P' from 'network' (CorrI2P-main/network.py)

Lose too many files in Kitti

it's best to generate it by yourself

Pre-trained model

Hello,

I have recently come across your paper and it is quite interesting.
I want to see if I can use it in a specific project. Is there a way to provide a pre-trained model (on KITTI for example) that I can use directly without having to train it again?

Thank you.

The code of visualization

Could you please provide the visual code of the matching connection graph in the paper?

Accuracy of model reproduction

Dear Siyu,
Thanks again for sharing your great job CorrI2P, I generated KITTI data according to DeepI2P and the way you mentioned without splicing multi-frame point cloud together, the point cloud number is 20480, and carried out experimental reproduction. After 25 epoches, training loss dropped from 4.78 to 3.34 and descriptor loss dropped from 3.95 to 3.17. The evaluation accuracy of the model is RTE 3.6402 + -49.7405 RRE 5.6616 + -19.2482. I am not sure if there are some problems with my dataset processing. It is too difficult to download the dataset you provided, and even if the download succeeds, many files are lost. The data processing script you uploaded looks no different from deepi2p's dataset , there is also multi-frame concatenation operation, so could you share your code for preprocessing Kitti? In addition, I see that the name of your point cloud file has been normalized. What does this normalization mean? I am looking forward to your reply.