Open-world semantic segmentation for Lidar Point Clouds

Official implementation of "Open-world semantic segmentation for Lidar Point Clouds", ECCV 2022. After saving the corresponding inference result files using this repository, please use semantic_kitti_api and nuScenes_api to evaluate the performance.

Installation

Requirements

• PyTorch >= 1.2
• yaml
• Cython
• torch-scatter
• nuScenes-devkit (optional for nuScenes)
• spconv (tested with spconv==1.2.1 and cuda==10.2)

Data Preparation

SemanticKITTI

./
├── 
├── ...
└── path_to_data_shown_in_config/
    ├──sequences
        ├── 00/           
        │   ├── velodyne/	
        |   |	├── 000000.bin
        |   |	├── 000001.bin
        |   |	└── ...
        │   └── labels/ 
        |       ├── 000000.label
        |       ├── 000001.label
        |       └── ...
        ├── 08/ # for validation
        ├── 11/ # 11-21 for testing
        └── 21/
	    └── ...

nuScenes

./
├── ...
├── v1.0-trainval
├── v1.0-test
├── samples
├── sweeps
├── maps
└── lidarseg/
    ├──v1.0-trainval/
    ├──v1.0-mini/
    ├──v1.0-test/
    ├──nuscenes_infos_train.pkl
    ├──nuscenes_infos_val.pkl
    ├──nuscenes_infos_test.pkl
└── panoptic/
    ├──v1.0-trainval/
    ├──v1.0-mini/
    ├──v1.0-test/

Checkpoints

We provide the checkpoints of open-set model and incremental learning model here: checkpoints

Open-set semantic segmentation

Training for SemanticKITTI

All scripts for SemanticKITTI dataset is in ./semantickitti_scripts.

MSP/Maxlogit method

./train_naive.sh

Upper bound

./train_upper.sh

RCF - Predictive Distribution Calibration

• Change the path of pretrained naive model in /config/semantickitti_ood_basic.yaml, line 63.

• Change the coefficient lamda_1 in /config/semantickitti_ood_basic.yaml, line 70.

• Change the dummy classifier number in /train_cylinder_asym_ood_basic.py, line 198.

./train_ood_basic.sh

RCF - Unknown Object Synthesis

• Change the path of pretrained naive model in /config/semantickitti_ood_final.yaml, line 63.

• Change lamda_1, lamda_2 in /config/semantickitti_ood_final.yaml, line 70, 71.

• Change the dummy classifier number in /train_cylinder_asym_ood_final.py, line 198.

./train_ood_final.sh

MC-Dropout

./train_dropout.sh

Evaluation for SemanticKITTI

We save the in-distribution prediction labels and uncertainty scores for every points in the validation set, and these files will be used to calculate the closed-set mIoU and open-set metrics including AUPR, AURPC, and FPR95.

MSP/Maxlogit

• Change the trained model path (Naive method) in /config/semantickitti.yaml, line 63.

• Change the saving path of in-distribution prediction results and uncertainty scores in val_cylinder_asym.py, line 112, 114, 116.

./val.sh

Upper bound

• Change the trained model path (Placeholder method) in /config/semantickitti.yaml, line 63.

• Change the saving path of in-distribution prediction results and uncertainty scores in val_cylinder_asym_upper.py, line 115, 117.

./val_upper.sh

RCF

• Change the trained model path (Placeholder method) in /config/semantickitti_ood_final.yaml, line 63.

• Change the saving path of in-distribution prediction results and uncertainty scores in val_cylinder_asym_ood.py, line 124, 125.

./val_ood.sh

MC-Dropout

./val_dropout.sh

Training for nuScenes

All scripts for nuScenes dataset are in ./nuScenes_scripts

MSP/Maxlogit method

./train_nusc_naive.sh

Upper bound

./train_nusc.sh

RCF - Predictive Distribution Calibration

• Change the path of pretrained naive model in /config/nuScenes_ood_basic.yaml, line 63.

• Change the coefficient lamda_1 in /config/nuScenes_ood_basic.yaml, line 70.

• Change the dummy classifier number in /train_cylinder_asym_nuscenes_ood_basic.py, line 197.

./train_nusc_ood_basic.sh

RCF - Unknown Object Synthesis

• Change the path of pretrained naive model in /config/nuScenes_ood_final.yaml, line 63.

• Change lamda_1, lamda_2 in /config/nuScenes_ood_final.yaml, line 70, 71.

• Change the dummy classifier number in /train_cylinder_asym_nuscenes_ood_final.py, line 197.

./train_nusc_ood_final.sh

MC-Dropout

./train_nusc_dropout.sh

Evaluation for nuScenes

MSP/Maxlogit

• Change the trained model path (Naive method) in /config/nuScenes.yaml, line 63.

• Change the saving path of in-distribution prediction results and uncertainty scores in val_cylinder_asym_nusc.py, line 112, 114, 116.

./val_nusc.sh

Upper bound

• Change the trained model path (Naive method) in /config/nuScenes.yaml, line 63.

• Change the saving path of in-distribution prediction results and uncertainty scores in val_cylinder_asym_nusc_upper.py, line 121, 123.

./val_nusc_upper.sh

RCF

• Change the trained model path (Placeholder method) in /config/nuScenes_ood_final.yaml, line 63.

• Change the saving path of in-distribution prediction results and uncertainty scores in val_cylinder_asym_nusc_ood.py, line 125, 126.

./val_nusc_ood.sh

MC-Dropout

./val_nusc_dropout.sh

Incremental learning

Training for SemanticKITTI

All scripts for SemanticKITTI dataset is in ./semantickitti_scripts.

First, use the trained base model to generate and save the pseudo labels of the training set:

./val_generate_incre_labels.sh

• Change the trained model path in /config/semantickitti_ood_generate_incre_labels.yaml, line 63.
• Change the save path of pseudo labels in val_cylinder_asym_generate_incre_labels.py, line 116.

Then, change the loading path of pseudo labels in /dataloader/pc_dataset.py, line 177.

Now, conduct incremental learing using pseudo labels:

./train_incre.sh

• Change the trained model path in /config/semantickitti_ood_incre.yaml, line 63.

Evaluation for SemanticKITTI

For validation set:

./val_incre.sh

For test set:

./test_incre.sh

• Change the collate_fn=collate_fn_BEV_val_test in /builder/data_builder.py, line 70.
• Change the save path in /dataloader/pc_dataset.py, line 95.
• Upload the generated files into the evalution server.

Training for nuScenes

All scripts for nuScenes dataset are in ./nuScenes_scripts.

For new class 1(barrier)

First, generate and save the pseudo labels of the training set:

./val_nusc_generate_incre_labels.sh

• Change the trained model path in /config/nuScenes_ood_generate_incre_labels.yaml, line 63.
• Change the save path of pseudo labels in val_cylinder_asym_nusc_generate_incre_labels.py, line 120.

Then, change the loading path of pseudo labels in /dataloader/pc_dataset.py, line 266.

Now, conduct incremental learing using pseudo labels:

./train_nusc_incre.sh

For new class 5(construction-vehicle), 8(traffic-cone), 9(trailer)

First, generate and save the pseudo labels of the training set:

./val_nusc_generate_incre_labels.sh

• Change the python file from val_cylinder_asym_nusc_generate_incre_labels.py into val_cylinder_asym_nusc_generate_incre_labels_1.py.
• Change the incremental class in val_cylinder_asym_nusc_generate_incre_labels_1.py, line 153.
• Change the trained model and save path similar with new class 1.

Then, change the loading path of pseudo labels in /dataloader/pc_dataset.py, line 266.

Now, conduct incremental learing using pseudo labels:

./train_nusc_incre.sh

• Change the trained model path in /config/nuScenes_ood_incre.yaml, line 63.

Evaluation for nuScenes

For validation set:

./val_incre.sh

For test set:

• Change the NuScenes version from nusc = NuScenes(version='v1.0-trainval', dataroot=data_path, verbose=True) to nusc = NuScenes(version='v1.0-test', dataroot=data_path, verbose=True)

./test_incre.sh

Then upload the generated files into the evaluation server.