search

zezeaaa / MVPGS

📑[ECCV'2024]MVPGS: Excavating Multi-view Priors for Gaussian Splatting from Sparse Input Views
https://zezeaaa.github.io/projects/MVPGS/
Other
32 stars 0 forks source link

Custom Dataset #2

Open hhhddddddd opened 1 month ago

hhhddddddd commented 1 month ago

hello,

I'm really interested in running your code, but I would like to use my custom dataset. My dataset consists of images of outdoor scenes captured using an iPad. I've already converted the video frames into individual images.

I'm quite new to the field of MVS and find some aspects challenging to grasp. Could you please provide detailed steps on how to run your code with my custom dataset? Any guidance would be greatly appreciated.

thank you!

zezeaaa commented 1 month ago

Hi,  Thank you for your interest! I’d be happy to add a custom dataset dataloader along with a corresponding guide, as I’m also curious about how MVPGS performs on custom datasets. However, I’d like to point out that MVPGS is designed for scenarios where camera poses are known (e.g., obtained through extrinsic calibration). In your case, maybe we can use COLMAP to solve for camera poses.  Additionally, MVPGS is originally designed for sparse-view input, where multi-view constraints are less abundant. If you use dense views as inputs, the improvement of performance might not be as significant due to more sufficient multi-view constraints.  I will work on providing a guide for custom datasets as soon as possible.

hhhddddddd commented 1 month ago

Thank you very much for your reply！Looking forward to the guidance on the custom dataset.

My dataset contains 17 training images, and the camera pose has been obtained through colmap.

I ran MVPGS on custom dataset and found that the initial point cloud generated by MVS has only one point. Is this because I haven't changed any settings? I haven't changed the "./mvs_modules/configs/config_mvsformer.json".Can you give me some advice?

thank you！

image

zezeaaa commented 3 weeks ago

Hello, I'm sorry for the late reply. MVS requires view overlapping to estimate depth. I have met such situations when I test MVPGS on MipNeRF360 dataset. I think we need to pick some views with good overlappings when applying MVS. Maybe I need more time to check whether it works because I'm working on another deadline recently. I will let you know if I have any updates.