search

bennyguo / instant-nsr-pl

Neural Surface reconstruction based on Instant-NGP. Efficient and customizable boilerplate for your research projects. Train NeuS in 10min!
MIT License
856 stars 84 forks source link

About Defining the Region of Interest in COLMAP format data #17

Open SAOHPRWHG opened 2 years ago

SAOHPRWHG commented 2 years ago

I follow the instruction and train the model on COLMAP format data but got this result. image I note that the preprocess in NeuS(https://github.com/Totoro97/NeuS/tree/main/preprocess_custom_data) has another step: image So how can I define the region of interest in this repo?

bennyguo commented 2 years ago

Hi! We automatically remove outliars in the sparse point cloud and normalize the cameras to lie in a unit sphere: https://github.com/bennyguo/instant-nsr-pl/blob/4f70db328827dee6596f1553df0177255f32a1c2/datasets/colmap.py#L39-L73 The region of interest of the object is by default set to [-0.5,0.5] as in the config file. If you want a tigher bounding box, you could visualize the normalized points like I do: https://github.com/bennyguo/instant-nsr-pl/blob/4f70db328827dee6596f1553df0177255f32a1c2/datasets/colmap.py#L183 and modify radius in the config file accordingly.

SAOHPRWHG commented 2 years ago

Thanks for your support very much! But I try to visualize and modify the radius, and the results do not improve. It seems it will always reconstruct a cube outside the aim object. image

SAOHPRWHG commented 1 year ago

Do you guys have some ideas about removing the cube outside the object? image

bennyguo commented 1 year ago

Hi, could you try our latest code and see if the floaters are reduced? See #26 for several improvements.