Have any one tried on custom dataset? I just tried the model and these are my result.

[wzy-99]

I tried several custom samples using mixRe10kDl3dv_512x512.ckpt. For intrinstics, I set fx=fy=1.0, cx=0.5, cy=0.5. For non-square images, I just center crop to process them.

1. Sora Data

   I use the images token from Figure 3 in the arxiv pdf file.

   The firsrt view is:

   

   The secord view is:

   

   The result looks nomal:

   

   But when I change the input sequence order:

   The firsrt view is:

   

   The secord view is:

   

   The reuslt looks more warpped:

   

2. toy vehicle

   

   

   

3. waymo

   

   

   

4. mount

   

   

   

The results are satisfactory in the center area and near the camera's source position, but they deteriorate with significant changes in the view.

(I also create a gradio app to test custom data at https://github.com/wzy-99/NoPoSplat-gradio )

[botaoye]

Hi, thank you for your great efforts. As pixelSplat suggests, the Gaussians at the boundaries are generally of lower quality because they are less constrained during training. As a result, these Gaussians may be large and cover other Gaussians as shown in your figures. Therefore, one suggestion is to remove the Gaussian corresponding to the image boundary (e.g. 14 pixels for a resolution of 512x512). I'm busy with other work right now and will update the demo when I have time.