search

yanx27 / PointASNL

PointASNL: Robust Point Clouds Processing using Nonlocal Neural Networks with Adaptive Sampling （CVPR 2020）
MIT License
261 stars 34 forks source link

about the evaluation performance on sequence08 of SemanticKITTI dataset #6

Closed amiltonwong closed 4 years ago

amiltonwong commented 4 years ago

Hi, @yanx27

I perform the evaluation on sequence08 of SemanticKITTI dataset by command: python test_semantic_kitti.py --model_path log/PointASNL_SemanticKITTI/latest_model.ckpt. It takes almost 4 days to finish the evaluation. Is this inference time normal or not? My system configuration is a Titan Xp GPU and CUDA 10.0 installed.

THX!

yanx27 commented 4 years ago

@amiltonwong Hi, the code python test_semantic_kitti.py uses small sliding window on raw point clouds, so it's relatively slow, you can try smaller voting number if you are time limited. Furthermore, we already update python train_semantic_kitti_grid.py and python test_semantic_kitti_grid.py, which is training and testing on grid sampled sub-point clouds and finally reproject prediction to raw point clouds, and it achieves better result of around 52% with several times efficiency. It will cost 1-2 days for training and 0.5 day for evaluation on sequence08 :)

amiltonwong commented 4 years ago

@yanx2, thanks a lot. I'll have a try in train_semantic_kitti_grid.py and test_semantic_kitti_grid.py

kxhit commented 4 years ago

@yanx27 Hi! Thanks for sharing your code. I notice that you achieved 62.3% IoU on SemanticKITTI benchmark. Is there any new paper with that result? I'm looking forward to knowing how the algorithm is designed. Thanks a lot!