torch

Torch implementation of occlusion detection and surface normal estimation

Data lists

The training and testing data lists can be found in ./data_list/. Specifically:

• sync_* are for occlusion detection.
• mp_* are for normal estimation on Matterport3D.
• scannet_* are for normal estimation on ScanNet.

Occlusion detection

Testing

Run

th main_test_bound_realsense.lua -test_model ../pre_train_model/bound.t7 -test_file ./data_list/realsense_list.txt -root_path ../data/realsense/

The result will be generated in ./result/bound_realsense_test_bound/. On ScanNet and Matterport3d, you may use main_test_bound_[scannet/matterport].lus to test on ./data_list/scannet_test_list_small.txt and ./data_list/mp_test_list_horizontal.txt.

Training

Run

th main_train_bound.lua -pretrain_file ../pre_train_model/sync.t7 -root_path ../data/pbrs_boundary/ -ps ./model/bound

The snapshots and models will be saved under ./model/.

Surface normal estimation

Testing

Run

th main_test_realsense.lua -test_model ../pre_train_model/normal_scannet.t7 -test_file ./data_list/realsense_list.txt -root_path ../data/realsense/

The result will be generated in ./result/normal_scannet_realsense_test/. On ScanNet and Matterport3d, you may use main_test_[scannet/matterport].lus to test on ./data_list/scannet_test_list_small.txt and ./data_list/mp_test_list_horizontal.txt.

Training

Run

 th main_train_matterport.lua -ps ./model/normal_matterport -use_render_normal_gt -root_path ../data/to_matterport/

The snapshots and models will be saved under ./model/. Noted that you need to download color image from official Matterport3D dataset in order to train the model. Same as before, you can use main_train_scannet.lua to train on ScanNet.