Fail to reproduce (Co3D dataset). Ques: Image resolution + Pretrained model

[WalterC9]

Hi, thank you very much for sharing your code. I have several questions regarding your work:

1. Could you specify the image resolution used during training for each scene in the Co3D dataset?

2. Would it be possible for you to share your pretrained Gaussians from the Co3D dataset? This would be very helpful for comparing qualitative results between different methods.

3. I encountered an out-of-memory issue while trying to reproduce the results on the Co3D dataset (though it works fine for the Tanks dataset). Your paper mentions that a single RTX 3090 is sufficient to train the model, so I expected the code to work on my RTX 4090. Did you face this issue during your training? If so, how did you resolve it?

Your help is greatly appreciated!

[OasisYang]

1. In most cases, we used the original resolution for scenes in the CO3D dataset.
2. Sure, I'm trying to find the rendered RGB images used in our paper and will upload them as soon as possible. To reproduce the results shown in our paper, please use the ZoeDepth as the depth estimator
3. I didn't face this problem for experiments on CO3D and Tanks&Temples, but I got this issue when running on some internet videos. This may caused by the inaccurate pose estimation as we didn't optimize the pose along with the optimization of the global gaussian model. Our follow-up work resolves this issue by a joint optimization on camera poses and 3D gaussian. Feel free to play with it.

[WalterC9]

Hi OasisYang,

Thank you for releasing the rendered images and depth maps for visual comparison — I greatly appreciate it!

However, I am having difficulty in (1) drawing poses visualization and (2) evaluating the rendered depth; as only the depth images are released instead of actual depth value.

I was wondering if it would be possible to also release the optimized Gaussians and poses. This would be very helpful to compare different model both quantitatively and quantitatively.

[Kevin-2017]

I have the same issue with reproducing the Co3d results. Has anyone else been able to reproduce the results? What is the hyperparameter used for co3d? Below is the command that I used.

python run_cf3dgs.py \
      -s "path" \
      -m "output/${scene}" \
      --mode train \
      --data_type co3d \
      --eval 

[WalterC9]

I have the same issue with reproducing the Co3d results. Has anyone else been able to reproduce the results? What is the hyperparameter used for co3d? Below is the command that I used.

python run_cf3dgs.py \
      -s "path" \
      -m "output/${scene}" \
      --mode train \
      --data_type co3d \
      --eval 

I found that using Zoe depth produce better results compared to others depth model. However, the best I got is still way worse than the ones reported in the paper.

On plant scene, I got 20.72 PSNR, while it is 29.69 in the paper. I evaluated the published results, but still the PSNR is only 23.00.

[mochan-shs]

The sequence number for 'plant scene' in the paper is "110 13051 23361"? I also have the same issue with reproducing the Co3d results.