train on PointPillars and PointRCNN

[TimGor1997]

Dear author,
May I ask if only 500/125 training samples will be used when retraining PointPillars and PointRCNN with 500frames and 125frames? Or 500/125 annotated samples plus 3712-500/125 unannotated samples?

[Cliu2]

Hi,

Thanks for the question. We use the 500/125 human annotated samples, plus the remaining 3712-500/125 samples (which are annotated by the MTrans auto labeler) to train the PointPillars/PointRCNN from scratch. So there are in total 500/125 human annotations plus 3712-500/125 neural-network-generated pseudo labels for training.

[TimGor1997]

Thank you for your reply~
But there is a question, how to get the PointPillars and PointRCNN only with 500/125 frames without the remaining 3712-500/125 samples?

[Cliu2]

Those rows are the results for PointPillars/PointRCNN trained with 500/125 human annotations only, no pseudo labels. We use the "500f"/"125f" to denote how many human annotations are required for these experiments. Please also check Sec. 5.2 paragraph 1 for details.

[TimGor1997]

Those rows are the results for PointPillars/PointRCNN trained with 500/125 human annotations only, no pseudo labels. We use the "500f"/"125f" to denote how many human annotation is required for these experiments. Please also check Sec. 5.2 paragraph 1 for details.

Thank you so much for your great patience and huge help.

Could you check if my understanding is correct?
The sencond line and the third line in PointPillars and PointRCNN is only train with 500/125 frames samples, total 500/125 training samples.
The fourth line is train wiht 500/125 human annotations + 2712-500/125 remaining samples, total 3762 training sampels.

[Cliu2]

Yes, that is correct. Thanks.

[TimGor1997]

Yes, that is correct. Thanks.
Thank you for the assistance you provided me, answering many of my questions. Looking forward to seeing your new research work!
All the Best!

[TimGor1997]

I'm terribly sorry, but I have another question to ask you
Why is the number of pseudo_labels out there only 3387 instead of 3769?
When you use OpenPCDet's code to compute AP3D, do you only compute these 3387?

[Cliu2]

We removed labels with too much truncation at here. The target objects have way too few LiDAR points to generate a good enough pseudo label, and therefore are omitted, resulting in slightly fewer pseudo labels than the human labels. Yes, only 3387 are used, during assessing the quality of pseudo labels.

[TimGor1997]

We removed labels with too much truncation at here. The target objects have way too few LiDAR points to generate a good enough pseudo label, and therefore are omitted, resulting in slightly fewer pseudo labels than the human labels. Yes, only 3387 are used, during assessing the quality of pseudo labels.

Thank you for your reply.
I found that when regenerating the train label, a total of 3387 pseudo labels were generated, excluding 500 manually marked labels for training, there were 2727 pseudo labels left. When training PointPillars/PointRCNN with 500 frames, 3712-500-2727=485 samples should we use manual labels or empty txt instead?

[Cliu2]

They are simply not supervised, no loss is calculated for those objects. Empty txt can be used.

[TimGor1997]

Copy that! Thank you very much for taking the time to answer my questions   TimGor ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: ***@***.***>; 发送时间: 2024年2月20日(星期二) 下午4:59 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [Cliu2/MTrans] train on PointPillars and PointRCNN (Issue #11) They are simply not supervised, no loss is calculated for those objects. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[TimGor1997]

Dear author I would like to ask, when retraining PointRCNN, does it only need to replace the pseudo label files in label_2, and does the OpenPCDet code need to be modified? Because I retrained MTrans (500), the result of training PointRCNN was much different from Table1, with a Hard difference of nearly 7% TimGor ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "Cliu2/MTrans" ***@***.***>; 发送时间: 2024年2月20日(星期二) 下午4:59 ***@***.***>; ***@***.******@***.***>; 主题: Re: [Cliu2/MTrans] train on PointPillars and PointRCNN (Issue #11) They are simply not supervised, no loss is calculated for those objects. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[TimGor1997]

[Cliu2]

In our experiments, we just replace the label_2 files with the generated pseudo labels. Albeit the results could vary from time to time due to the randomness of different environments, the large gap is still strange. Have you checked the mIoU of generated pseudo labels against the ground truth labels? Does it match with the paper?