Open fariba87 opened 1 month ago
This is wired. Did you check if the camera poses are correct ? I have provided a colmap extractor in the convert.py for your reference.
This is wired. Did you check if the camera poses are correct ? I have provided a colmap extractor in the convert.py for your reference.
yes. in convert.py camera model is set as "OPENCV" . i used GUI version of colmap to create sparse folder both with "pinhole" and "opencv" model. but still same problem. i tried to evaluate render.py up to the point loss is not nan. but the final mesh is wrong(even though point cloud created during training seems right ]
How about the rendering results?
hello. thank you for your great work. i want to do 3d reconstruction on a small object, which i captured around 40 images while it is circling on a turntable. i used pinhole camera model in colmap GUI to extract camera parameters. the images' resolution is ~4000*3000 and i consider to use principle point in the center of image. i have some questions 1) should i use undistorted images for training? 2) for training i use python train.py -s mydatasetpath -r 2 --depth_ratio 1 --lambda_dist 100 --lambda_normal 100 but after some iteraration(~100) loss goes to nan. what should be the reason? 3) sometime by resizing images( 16 times smaller) loss will not be nan but it is not decreasing and just fluctuating around a number. how can i change the parameters to have decreasing loss? 4) besides, the number of points is decreasing!