Question about MatrixCity Street dataset experiments

[BitCalSaul]

Hello,

I read your paper and found the results on the MatrixCity Street dataset very impressive — especially reaching a PSNR of 25.

I would like to ask a few clarifications:

For the Street dataset, did you use small_city/road_down_vertical/outside, horizon, or the entire set (all)?

How many iterations were used for training?

Would it be possible to provide the pretrained model file?

Thank you very much for your great work!

[SunshineWYC]

Thank you for your interest. For the street dataset, we use all available angles (a total of 4,075 training views). For the hyperparameter iterations, we set it to 30,000 to ensure a fair comparison with other methods.

The result for the mc_street is available at the following link.

Hope this is useful.

[BitCalSaul]

Thanks very much for your timely reply. I just noticed that in the MatrixCity street dataset, there are over 28,000 training views. Could you clarify which subset you used exclusively for training?

[BitCalSaul]

@SunshineWYC

I have one more question: for the Mill19 and UrbanScene3D datasets (the ones reported in Table II of your paper), would it be possible for you to also share the pretrained model checkpoints?

It would be very helpful for reproducing your results and making fair comparisons.

Thanks again for your great work!

[SunshineWYC]

For the MatrixCity street dataset, the image subset for Block_A street views was constructed following the existing methods.

The results for pre-trained models on other scenes will be available in the coming days. Thank you for your patience.

[SunshineWYC]

Now we have uploaded the others optimized scene models, you can find the ply file here.

[BitCalSaul]

Hi @SunshineWYC,

Many thanks for your very prompt and helpful replies — really appreciate the detailed clarifications and for sharing the pretrained models so quickly!

I just have one follow-up question: could you kindly let me know where I can find more details about how the Block_A street dataset subset was constructed? It would be very helpful for us to better reproduce your reported results.

Thanks again for your great work!

[BitCalSaul]

Additionally, I noticed a possible issue with the uploaded .ply file (points3D_streetAblockgs.ply). When I load it using a simple script (python read_ply_points.py ...), the reported coordinate ranges are:

File: point_cloud_merged.ply
Number of points: 5989675
Coordinate ranges:
  X: [nan, nan]
  Y: [nan, nan]
  Z: [nan, nan]

It seems the coordinates are all NaN. Could you kindly check whether the file was uploaded correctly?

[BitCalSaul]

Hi @SunshineWYC, thanks again for sharing the pretrained models — really appreciate it!

I tried evaluating the MatrixCity aerial pretrained .ply (≈37M Gaussians) on the official test split (741 images). I loaded all attributes (xyz, scale, rot, feature, opacity). However, I only obtain ~24.x PSNR on the test set. May you kindly give me any hint?

[SunshineWYC]

It's recommended to verify the evaluation scripts to make sure aligned with the repo. In addition, if rendered images are saved as jpg format, the image quality degradation is non-negligible.

…

---Original--- From: ***@***.***> Date: Tue, Sep 30, 2025 23:45 PM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [SunshineWYC/BlockGaussian] Question about MatrixCity Streetdataset experiments (Issue #29) BitCalSaul left a comment (SunshineWYC/BlockGaussian#29) Hi @SunshineWYC, thanks again for sharing the pretrained models — really appreciate it! I tried evaluating the MatrixCity aerial pretrained .ply (≈37M Gaussians) on the official test split (741 images). I loaded all attributes (xyz, scale, rot, feature, opacity). However, I only obtain ~24.x PSNR on the test set. May you kindly give me any hint? Cursor.2025-09-30.10.41.29.png (view on web) — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>

[SunshineWYC]

For matrixcity_street scene's view, the initial script is launched in CityGaussian repo. Hope it is helpful.

As for our pretrained model points3D_streetAblockgs.ply, thanks for your attention, we will verify it as soon as possible!

[BitCalSaul]

Hi @SunshineWYC, thanks a lot for the clarification!
I directly compared the rendered images in memory (not saved as JPG), so image compression should not be the cause. I’ll carefully check the evaluation pipeline, especially around this part of the script, to make sure my evaluation is fully aligned with your repo.