Shaoyu Chen
Shaoyu Chen
Please set MAX_PRIMARY_NUM to a larger number (refer to hierarchical_aggregation.cu Line 8). And then re-compile the external C++ and CUDA ops.
We plan to release the code before the main conference of ICCV 2021. Currently we still need some time to refactor the code, especially the CUDA implementation, making it more...
If cfg.use_coords is True, the input_channel is 6 (3 for r, g, b and 3 for x, y, z). Please refer to Line 132 in hais.py.
Thanks for your attention to our work. It takes 4 days to train HAIS for 500 epochs on a single Titan X GPU. Less epochs also achieve comparable performance. And...
For batch size = 4, ~12G is OK.
You need to modify the definition of class_numpoint_mean_dict (Line 7 in hierarchical_aggregation.cpp) and class_radius_mean (Line 55 in hierarchical_aggregation.cu), according to the statistics of your data.
Thanks for benchmarking HAIS on STPLS3D! I have updated the information about STPLS3D on our project page. Most point cloud datasets for instance segmentation (ScanNet, S3DIS, SceneNN, etc.) only focus...
It's OK. And I have already updated the result.
Hi. Thanks for your attention. The code and model of S3DIS are not available currently. Experiments on S3DIS are not compatible with the released code and reorganizing them is really...