help！！！！！！！！！！！！！！！！！！！！！！！！！！！！！！！！！

[honghu404]

I have now shot a video orbiting an object, split it into many images, and plan to build a dataset for training on your method next. Besides these images, what parameters are required for each image? Could you please inform me in detail? Thank you.

[zhizdev]

Hey! Thanks for checking out our code base. We used fixed cameras views (rendered from blender) for training and find that if the camera views are from an orbiting video, multi-view generation methods may struggle.

To render objects on blender with our fixed views, please visit: #10 (comment)

To use on arbitrary cameras please checkout MVSplat360 https://github.com/donydchen/mvsplat360

[honghu404]

Hey! Thanks for checking out our code base. We used fixed cameras views (rendered from blender) for training and find that if the camera views are from an orbiting video, multi-view generation methods may struggle.

To render objects on blender with our fixed views, please visit: #10 (comment)

To use on arbitrary cameras please checkout MVSplat360 https://github.com/donydchen/mvsplat360

Existing multi-view image generation methods mainly target anime characters, simple objects, or human faces, and all involve fixed-position objects with 360° surrounding multi-view image capture. The task I need to accomplish is multi-view generation for buildings. What I can obtain is only a video shot by surrounding certain buildings, but the distance from the camera to the object may vary slightly at each position. I can split this video into individual images, and use COLMAP to obtain the relative pose relationships of each image, or some sparse point clouds, etc. I would like to ask for your advice: next, I want to build a dataset using the data I can obtain for building multi-view generation. The trained model should take as input an image of the building at a certain moment, its pose, or other accessible inputs, and output multi-view images of the building from different perspectives. Do you have any suggestions regarding dataset construction and method implementation? Thank you very much!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

[zhizdev]

I would start with looking into mvsplat360 or feed forward methods like VGGT