about --max_time in train_eval.sh

[SleepingPurin]

In section 4.1 of the paper on arXiv, it mentions timestamps ranging from 0 to 0.7 as the training set. However, among the four datasets (Dynamic Objects, Dynamic Indoor Scenes, Dynamic Multiparts, and NVIDIA Dynamic Scenes) downloaded from the link provided in this repository, the maximum timestamp in the training part is 0.75. In the train_eval.sh file, the value of --max_time is not 0.75, but rather 0.7 or 0.65. Could you please clarify under which settings the experimental data presented in the paper (Table 10, 11, 12, 13) were obtained?
I trained the fan scene using "python train_gui.py -s /path/to/data/ -m $output/$exp --max_time 0.7 --light" and the chessboard scene using "python train_gui.py -s /path/to/data/ -m $output/$exp --max_time 0.65". These two commands have the same parameters as those in train_eval.sh. However, the MAEPSNR of the extrapolated future frames obtained did not match the results presented in the paper, being approximately 3 lower. Is this related to the --max_time parameter?

[Jinxi-Li]

Hi,

Thanks for your interest!

The timestamp in the training part is 0.75, that is right. The --max_time in the code is the max_time for the deformation field. When we were drafting the paper, I was using 0.7 as --max_time. But I cleaned the code and rearchitectured some of the content and tuned the hyperparameter a little bit, so it is impossible to get the exact same number as the paper. But I believe the code version should be nearly consistent to the paper. And the average performance should be slightly better.

I have checked my local results using the same code before I upload it, the MAEPSNR for fan is 35.282 (34.964 in paper) and for chessboard is 35.011 (35.271 in paper), which are comparably consistent I think.

I'll pull the code to my local machine and rerun it, to check if I can meet the same problem as you. I'll respond later once I have the results.

Once again, I need to emphasize, the --max_size argument only influence the deformation field. No matter how you choose it, the dataset is the same.

[Jinxi-Li]

Hi,

I still get the consistent results as this. I am not sure why the results are not right in your end. Can you provide more informaiton?

Hi,

Thanks for your interest!

The timestamp in the training part is 0.75, that is right. The --max_time in the code is the max_time for the deformation field. When we were drafting the paper, I was using 0.7 as --max_time. But I cleaned the code and rearchitectured some of the content and tuned the hyperparameter a little bit, so it is impossible to get the exact same number as the paper. But I believe the code version should be nearly consistent to the paper. And the average performance should be slightly better.

I have checked my local results using the same code before I upload it, the MAEPSNR for fan is 35.282 (34.964 in paper) and for chessboard is 35.011 (35.271 in paper), which are comparably consistent I think.

I'll pull the code to my local machine and rerun it, to check if I can meet the same problem as you. I'll respond later once I have the results.

Once again, I need to emphasize, the --max_size argument only influence the deformation field. No matter how you choose it, the dataset is the same.

[SleepingPurin]

Thank you very much for your patient reply.
I cloned the repository and set up the environment following the instructions provided. I also downloaded the datasets as required. I ran the experiments on the AutoDL platform using rented computing resources. I tested the Fan and Chessboard scenes, and obtained the following MAE-PSNR results for future-frame extrapolation:

The MAE-PSNR of the extrapolated future frames is only around 32 dB. I am using a rented vGPU (32GB) on the AutoDL platform instead of an NVIDIA RTX 3090 GPU — could this difference in hardware lead to the performance drop?

[Jinxi-Li]

The Chessboard look right, it's 35.0099, but the fan.py looks not right. May I ask where you download the dataset? I recommend the hugginface version for the exact reproduction.

[SleepingPurin]

The Chessboard is 32.999,not right.I download the dataset from the link provided in this repository.

[Jinxi-Li]

I see! It's strange. Can you try another scene to see if there is also performance drop? And can you check the qualitative results to see whether the rendered results and ground truth are matching?

[SleepingPurin]

I try the whale scene.The result is close to the paper.

[SleepingPurin]

I also re-rented an RTX 3090 with 24GB memory, reconfigured the environment, and reran the experiments on the entire Dynamic Object dataset using exactly the same commands as in the repository (including all parameters such as --max_time). The results are shown below. All other results look correct, except for the falling ball sequence. I suspect this might be related to setting --max_time to 0.5.