megvii-research / far3d Goto Github PK

I observe that some frames do not have all cameras synchronized during inference stage. And the validation fails if get_test_data() returns None. How to evaluate the whole frames for 23547 validation set?

请问nori2这个库是怎么装的，好像不是pypi里的

老师，执行python tools/create_infos_av2/create_av2_infos.py命令，无法生成av2_train_infos.pkl，出现了如下问题：

请老师教我怎么解决这个问题，感谢，盼复

感谢您开源，期待好久了，日后会提供nuscenes训练的config吗 或者方便发我一份吗 我的邮箱是[email protected]

Can you describe the latency of this model? and how about the speed compared with previous methods, e.g., Sparse4D?

can you share pip env like this https://github.com/megvii-research/Far3D/blob/main/py38.yaml in torch==1.12 ,becase flash-attention didn't support torch 1.11 below @dc3671 @zsc

Thank you for your great work and contributions to the open-source community.
Could you please share the NuScenes configuration with me? My email is [email protected].
Thanks again.

Thanks for your great work and open source!
Can not find the nuscenes config in the repo. Could you please share it with me? Your assistance is much appreciated! You can reach me via email: [email protected]

@xhjiang-pixel
I see that there are many people requesting for config file on NuScenes dataset. Is it possible to upload the config file to the repository so that people don't have to expose their email ids?

This will also create less trouble for you as responding to every single request is time consuming.

Hi,
I managed to initiate training with Nuscenes but apparently if Batch size > 1, there's a 'NotImplementedError'

        if B == 1:
            new_reference_points = coords3d.unsqueeze(0)    # (B, M, 3)
        else:
            print("Batch B:", B)
            raise NotImplementedError

in models/de

by any chance do you have an update for this part or do you guys still train for Batch size = 1?

Thanks!

hi ,can you share with me the nuscenes config? many thanks! email : [email protected]

Thanks for your great work and open source!
Can I have a reference config for NuScenes data ([email protected])?
Thanks!

Hello, can you provide the configuration file of nuscenes? After I modified it myself, the running result was not good. And epoch is 24.
mAP: 0.3832
mATE: 0.6107
mASE: 0.7045
mAOE: 1.5487
mAVE: 0.3040
mAAE: 1.0000
NDS: 0.3297
Eval time: 168.7s

Per-class results:
Object Class AP ATE ASE AOE AVE AAE
car 0.589 0.454 0.748 1.585 0.272 1.000
truck 0.307 0.684 0.792 1.579 0.261 1.000
bus 0.355 0.753 0.855 1.582 0.641 1.000
trailer 0.127 0.858 0.810 1.555 0.195 1.000
construction_vehicle 0.100 0.946 0.707 1.492 0.125 1.000
pedestrian 0.446 0.634 0.331 1.593 0.371 1.000
motorcycle 0.368 0.559 0.796 1.519 0.405 1.000
bicycle 0.351 0.456 0.809 1.724 0.163 1.000
traffic_cone 0.636 0.346 0.309 nan nan nan
barrier 0.554 0.417 0.887 1.309 nan nan

Thanks for sharing your work, it's excellent research.
Additionally, I was wondering if it would be possible for me to request a config of the Nuscenes data for reference?
I'm looking forward to your response.

How to adjust resize_lim based on input_size and final_dim？

同 #7 和 #9，我使用您提供的create_av2_info文件生成100ms间隔的av2数据索引后，在验证集中仍存在cam_info为None的帧，这会导致在测试时get_data_info()的返回值为None并引发后续报错。想请教下您是如何解决这个问题的。

Thank you for opening source of this great work. Do you mind to share one of your pretrained model for eval usage which could go with the config file provided? Thanks

I tried training Far3d on NuScenes.

It fails on the line
gt_ins_depth = [img_meta['ins_depthmap'] for img_meta in img_metas]
due to missing ins_depthmap.

Any hints how I can get that with NuScenes?

Thank you!

请问，这个UPath无法导入怎么解决？在网上没有查到解决版本，pip install的upath版本是1.0

Hello
First of all, thank you for sharing the great work.

I've noticed that the code used the nori2 package to construct dataloader. The nori2 package fetches data from Amazon S3.

I believe there might be another code snippet that handles dataloading from local storages without using nori2 package.
If that is the case, could you kindly make that code available?
Thank you for your consideration 🙏

您好，感谢开源，
现在生成了nuscene的pkl, 我看config里面是argoverse的，我改如何跑nuscene数据，方便解答一下，十分感谢！

这个 refile 是什么包？怎么安装？

Hello, Can you send me the Config for Nuscene dataset? My email is: [email protected]. Thank you!

I have used your code (python tools/create_infos_av2/create_av2_infos.py) to generate av2_val_infos_mini.pkl. However, I noticed that several cameras are empty, such as
'ring_front_left': None,
'ring_front_center': None,
'ring_rear_right': None
in 'scene_id': '02678d04-cc9f-3148-9f95-1ba66347dff9'.
Could you please provide me with a genuine pkl file? Thank you.

I am wondering if it is possible to export the model in ONNX format. Is it even possible? Is there any unsupported operation?
Has anyone tried it before?
Thanks!

Thank you for your great research and share your work!
Could i also receive nuScenes config?? （[email protected]）
Thanks

感谢您杰出的工作！
想参考下您在nuscenes数据集上的config，不知是否可以分享，我的邮箱是 [email protected]，非常感谢！

以及我在生成argoverse 2 info文件时，会有数行如下关于数据集图像缺失的提示：
No corresponding ring image at 315969904359876000: 27.5 > 25.0 ms
No corresponding ring image at 315969904359876000: 40.1 > 25.0 ms
No corresponding ring image at 315969904359876000: 32.6 > 25.0 ms
No corresponding ring image at 315966060559935000: 27.5 > 25.0 ms
No corresponding ring image at 315969614859771000: 27.7 > 25.0 ms
No corresponding ring image at 315969614859771000: 32.7 > 25.0 ms
......
想请教下这个是av2数据集的正常现象吗？