YoYo000
Custom unstructured data
Howdy!,
Thank you and team for making MVSNet code public. When running MVSNet using extrinsic and intrinsic obtained from colmap the output yields a very incorrect output and does not duplicate results seen using same image data . Blue points are generated with default extrinsic info in cameras folder and yellow points are generated from camera data obtained with conversion using colmap2mvsnet.py supplied within MVSNet. Files attached of both below. I'm exporting camera.txt from colmap using 'Export model as text' could that be the problem ? Any insight would be greatly appreciated
Hello, could you check whether the depth maps are as expected? (You can use the provided visualize.py). The depth maps are 0000*_init.pfm.
I dont see a way to send a direct message but can someone point me to a working example of this library. All training examples fail with no changes to the code due to path issues (after it finds the data and starts running) or out of memory issues as seen here
Resource exhausted: OOM when allocating tensor with shape[1,32,128,144,192] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
I have a RTX 2070 video card so it should have plenty of memory for this.
Ill delete this comment later as it isnt related to the question but I couldnt find another place to ask for help on running this code for the first time. Thanks for pointing me in the right direction!
depth maps look quite garbled . It happened to be the scale of the transformation matrix generated from col map , I've increased to 400 - 900 range and now working on scan1 dataset but generating strange results with tanks and temples dataset using provided pretrained model with limitations to 49 points when fusing, any ideas as to what may be going on?
@BingEdison The scale of the depth range doesn't matter. Make sure you set correct max_d and interval_scale. For the parameters for mvsnet, max_d * interval_scale should roughly equal to the max_d in the camera files. For example, if max_d=192 in the camera files, you can have max_d=256 and interval_scale=0.8
thanks!!! @jzhangbs only adjusted max_d and interval scale as suggested and now working with starting to work and with other captures also. Any tips on getting a better result from the head capture seen attached?
