problems running test on SUNCG

[yinlingluo]

Hi! I am running test on SUNCG test data (followed the instructions in README/Inference on SUNCG dataset). Q1: Should in Inference on SUNCG dataset/3, should we modify experiments/cfgs/ScanNet/benchmark.yml or experiments/cfgs/SUNCG/rpn_class_mask_5.yml?

Then I modified both and try to run the test. Q2: Those images files cannot be accessed since scene_name(defined in dataset.py/151) refers to first chunk(before"0") of scene_name, while all those real scene_names (as in suncg_test_data/images; suncg_test_data/scenes downloaded from given link) include the "0".

Then I tried either modify the directory names for those scenes in suncg_test_data/images or just change the definition of scene_name to include the "0". Q3: But then, after "starting test on whole scan..." I got: error: no invalid depth error: no invalid depth error: no invalid depth [path to the scene file]:(invalid sample: no valid projection) and this seems further caused "CUDA error:out of memory" issue

Is there anything I can do to solve this? Thanks!

[Sekunde]

Q1: experiments/cfgs/SUNCG/rpn_class_mask_5.yml Q2: can you change the folder name from suncg_test_data/images to suncg_test_data/images_augmented, since i just noticed the image folder name is hardcoded in the data loader, ref: https://github.com/Sekunde/3D-SIS/blob/master/lib/datasets/dataset.py#L144-L149 then the images should be found accordingly Q3: some suncg scenes are really big, thus we inference with a 24 GB GPU. You can change to use CPU by set parameters in https://github.com/Sekunde/3D-SIS/blob/master/lib/utils/config.py#L123-L124 If using more than 400 images and the volume is bigger than that threshold, it will automatically choose to inference on CPU. ref: https://github.com/Sekunde/3D-SIS/blob/master/lib/model/trainval.py#L664-L667 and https://github.com/Sekunde/3D-SIS/blob/master/lib/nets/network.py#L210-L214