The experimental results

[pg131413]

Hi, thank you for sharing your excellent work and code.

I am currently trying to reproduce the experimental results reported in your paper. I have a question regarding the evaluation metrics in the results table.

Could you please clarify whether the reported values are the average results over multiple runs, or the best results obtained from a single run/checkpoint?

If they are averaged results, could you also let me know how many runs were used and whether different random seeds were applied?

Thank you very much for your time and help!

[Ouasfi]

Hi @pg131413 we used a single run for all the methods I reproduced using the same seed (randomly generated and fixed). I think all the seeds are fixed in the code and you should be able to reproduce the exact numbers of the paper. The checkpoints are selected using the chamfer distance wrt the sparse input pointcloud.

[pg131413]

Hi, thank you very much for your clarification. I have one more question about the reported results on ShapeNet. Since ShapeNet contains multiple object categories, I would like to know how the evaluation metrics in the paper were computed across different categories. Are the numbers reported in the table averaged over all categories / all test samples, or are they the best results selected from individual categories? In other words, for each metric reported in the paper, does it represent the overall performance on the whole ShapeNet test set, or the best performance among different ShapeNet categories? This information would be very helpful for reproducing and comparing the results correctly. Thank you again! 发自我的iPhone

…

------------------ Original ------------------ From: Ouasfi ***@***.***> Date: Fri,May 22,2026 10:52 AM To: Ouasfi/SparseOcc ***@***.***> Cc: pg131413 ***@***.***>, Mention ***@***.***> Subject: Re: [Ouasfi/SparseOcc] The experimental results (Issue #4)

[Ouasfi]

For ShapeNet we followed the setup of Neural Splines as mentionned in the paper. If I remember correctly it uses the first N (i think N =20 ) test shapes of each classe and computes the average performance.

[pg131413]

I cannot reproduce the experimental results reported in your paper. Could you please release the detailed training configuration parameters and evaluation specifics for each dataset so that I can further attempt the reproduction? Thank you

[Ouasfi]

Hi @pg131413, which dataset are you trying to reproduce. I think in the paper i used the same config for all the datasets except 3Dscenes I think. Can you give more details about what you couldn't reproduce ?

[pg131413]

We select 20 files with the optimal weights from 100 pieces of human body data (000 to 099), and the experimental results reproduced from these files are significantly higher than the metrics reported in the original paper. Please elaborate on the evaluation details of the human body dataset. 发自我的iPhone

…

------------------ Original ------------------ From: Ouasfi ***@***.***> Date: Tue,Jun 2,2026 10:52 AM To: Ouasfi/SparseOcc ***@***.***> Cc: pg131413 ***@***.***>, Mention ***@***.***> Subject: Re: [Ouasfi/SparseOcc] The experimental results (Issue #4) Ouasfi left a comment (Ouasfi/SparseOcc#4) Hi @pg131413, which dataset are you trying to reproduce. I think in the paper i used the same config for all the datasets except 3Dscenes I think. Can you give more details about what you failed to reproduce ? — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS and Android. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[Ouasfi]

So you are trying to reproduce the Faust results. First can you please check this issue and make sure that you are able to obtain the same numbers ?

[pg131413]

#1 (comment)  Your dataset only contains a single sample, which leads to the reproduced metrics such as CD-L1 of 0.002437 and normal consistency of 0.96. However, FAUST dataset includes 100 point cloud samples in total. I wonder whether the reported metrics are the best results of a single sample or the average value across all samples. 发自我的iPhone

…

------------------ Original ------------------ From: Ouasfi ***@***.***> Date: Tue,Jun 2,2026 11:53 AM To: Ouasfi/SparseOcc ***@***.***> Cc: pg131413 ***@***.***>, Mention ***@***.***> Subject: Re: [Ouasfi/SparseOcc] The experimental results (Issue #4)

[Ouasfi]

I added this example for reproduction purposes. The full Faust dataset can be found here.

[pg131413]

I would like to ask whether the original evaluation metrics of the FAUST human dataset you reported, including CD1 0.26, CD2 0.002, Nc 0.952 and FS 0.974, are the average results of all samples in the dataset or the best single-sample result. 发自我的iPhone

…

------------------ Original ------------------ From: Ouasfi ***@***.***> Date: Tue,Jun 2,2026 0:13 PM To: Ouasfi/SparseOcc ***@***.***> Cc: pg131413 ***@***.***>, Mention ***@***.***> Subject: Re: [Ouasfi/SparseOcc] The experimental results (Issue #4) Ouasfi left a comment (Ouasfi/SparseOcc#4) I added this example for reproduction purposes. The full Faust dataset can be found here. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS and Android. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[Ouasfi]

The numbers in the paper are the average over all samples in the dataset.

[pg131413]

However, I cannot reproduce the results from the original paper by following the official configuration parameters. Could you please disclose all detailed training and evaluation procedures for the FAUST dataset in full? 发自我的iPhone

…

------------------ Original ------------------ From: Ouasfi ***@***.***> Date: Tue,Jun 2,2026 0:20 PM To: Ouasfi/SparseOcc ***@***.***> Cc: pg131413 ***@***.***>, Mention ***@***.***> Subject: Re: [Ouasfi/SparseOcc] The experimental results (Issue #4) Ouasfi left a comment (Ouasfi/SparseOcc#4) The numbers in the paper are the average over all samples in the dataset. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS and Android. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[Ouasfi]

Ok so you are able to reproduce the results of this example but not the other samples ? What information you think is missing? if you are able to reproduce this example numbers you should be able to reproduce the rest using the same configs for training and eval.

[pg131413]

Please answer my question carefully and fully understand what I mean. I want to ask whether the metrics CD1 0.260, CD2 0.002 and Nc 0.952 shown in Table 2 of the paper are the evaluation results of a single sample or the average results of the entire FAUST dataset. Please read my words thoroughly and do not give irrelevant replies. 发自我的iPhone

…

------------------ Original ------------------ From: Ouasfi ***@***.***> Date: Tue,Jun 2,2026 0:50 PM To: Ouasfi/SparseOcc ***@***.***> Cc: pg131413 ***@***.***>, Mention ***@***.***> Subject: Re: [Ouasfi/SparseOcc] The experimental results (Issue #4) Ouasfi left a comment (Ouasfi/SparseOcc#4) Ok so you are able to reproduce the results of this example but not the other samples ? What information you think is missing? if you are able to reproduce this example numbers you should be able to reproduce the rest using the same configs for training and eval. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS and Android. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[Ouasfi]

As I already explained, the numbers in the paper are the average results over the 100 samples from Faust dataset. You mentioned that you are not able to reproduce these numbers. So I wanted to check first if you are able to reproduce the result on one shape for which the sparse pointcloud is provided in the repo and that was reproduced by other users in the issue I mentioned. If you are able to reproduce this example you should be able to do the same on the 99 other using the same config and evaluation.