This is the repo for point clouds normalization. It currently supports the rotation regression based on PointNet.
This project is implemented based on PyTorch 1.5. Activate the training environment with the following command for the corresponding servers.
For SH40, run the following command:
source env40.sh
For SH36, run the following command:
source env36.sh
Some training parameters are set as follows by default:
| Param | Value |
|---|---|
| batch_size | 64 |
| epochs | 30 |
| LR | 1e-3 |
| npoints | 30000 |
| width_threshold | [-20, 20] |
Modify the specific training parameter with training command. To start training process, run the following command:
sh script/slurm_training.sh train TRAINING_NODE GPU_NUM --model_tag CUSTOM_TAG
An example of training on SH36:
sh script/slurm_training.sh train ad_lidar 8 --model_tag baseline
The trained model parameters are saved in ./output/model_tag/ckpt. If you want to train a model with different depth or width threshold, add the corresponding paramter like:
sh script/slurm_training.sh train ad_lidar 8 --model_tag baseline --depth_threshold --width_threshold
To obtain the final rotation result for a point sequence, run the following command:
sh script/slurm_correct.sh correct TRAINING_NODE GPU_NUM --model_tag CUSTOM_TAG --pretrained_model MODEL_PATH --pkl_file PKL_FILE_PATH
If you want to have a detailed result about the abnormal predictions, you can set an error threshold to select the corresponding point sequence. To be specific, run the following command:
sh script/slurm_correct.sh correct TRAINING_NODE GPU_NUM --model_tag CUSTOM_TAG --pretrained_model MODEL_PATH --pkl_file PKL_FILE_PATH --to_file --max_threshold MAX_THRESHOLD --min_threshold MIN_THRESHOLD
A file named abnormal.txt wil be saved in ./output/model_tag/ with selected point sequence index and its predicted rotation which is greater than the MAX_THRESHOLD or less than the MIN_THRESHOLD.
An avaiable pretrained model is saved in output/model_best/best.pth