Skip to content

Latest commit

 

History

History
112 lines (90 loc) · 5.27 KB

README.md

File metadata and controls

112 lines (90 loc) · 5.27 KB
Raw

RELLIS-3D Benchmarks

The HRNet, SalsaNext and KPConv can use environment file requirement.txt. GSCNN need use file gscnn_requirement.txt.

Image Semantic Segmenation

Note: New script for evaluate the results is available for point cloud and image

HRNet+OCR

The HRNext+OCR is a fork from https://github.com/HRNet/HRNet-Semantic-Segmentation/tree/HRNet-OCR

To evaluate the dataset:

cd /path/to/code/benchmarks/HRNet-Semantic-Segmentation-HRNet-OCR
export PYTHONPATH=/path/to/code/benchmarks/HRNet-Semantic-Segmentation-HRNet-OCR/:$PYTHONPATH
python tools/test.py --cfg experiments/rellis/seg_hrnet_ocr_w48_train_512x1024_sgd_lr1e-3_wd5e-4_bs_12_epoch484.yaml \
                     DATASET.TEST_SET val.lst \
                     OUTPUT_DIR /path/for/save/prediction \
                     TEST.MODEL_FILE /path/to/code/benchmarks/HRNet-Semantic-Segmentation-HRNet-OCR/output/rellis/seg_hrnet_ocr_w48_train_512x1024_sgd_lr1e-3_wd5e-4_bs_12_epoch484/best.pth

Add dataset path to ROOT in experiments/rellis/seg_hrnet_ocr_w48_train_512x1024_sgd_lr1e-3_wd5e-4_bs_12_epoch484.yaml

The models are initialized by the weights pretrained on the ImageNet. You can download the pretrained models from onedrive or https://github.com/HRNet/HRNet-Image-Classification.

Note: the pre-trained model was updated on June 8th 2021

Dowload pre-trained model (Download 751MB)

To retrain the HRNet on RELLIS-3D:

export PYTHONPATH=/path/to/code/benchmarks/HRNet-Semantic-Segmentation-HRNet-OCR/:$PYTHONPATH
echo $PYTHONPATH
PY_CMD="python -m torch.distributed.launch --nproc_per_node=2"
$PY_CMD tools/train.py --cfg experiments/rellis/seg_hrnet_ocr_w48_train_512x1024_sgd_lr1e-3_wd5e-4_bs_12_epoch484.yaml

Add dataset path to ROOT in experiments/rellis/seg_hrnet_ocr_w48_train_512x1024_sgd_lr1e-3_wd5e-4_bs_12_epoch484.yaml

GSCNN

The GSCNN is a fork from https://github.com/nv-tlabs/GSCNN

To evaluate the dataset:

cd /path/to/code/benchmarks/GSCNN-master
export PYTHONPATH=/path/to/code/benchmarks/GSCNN-master/:$PYTHONPATH
python train.py --dataset rellis --bs_mult 3 --lr 0.001 --exp final \
                --checkpoint_path /path/to/pre-trained/chk_file \
                --mode test \
                --test_sv_path /path/for/save/prediction

Add dataset path to __C.DATASET.RELLIS_DIR in benchmarks/GSCNN-master/config.py Dowload pre-trained model (Download 1GB)

To retrain the GSCNN on RELLIS-3D:

export PYTHONPATH=/path/to/code/benchmarks/GSCNN-master/:$PYTHONPATH
python train.py --dataset rellis --bs_mult 3 --lr 0.001 --exp final

Add dataset path to __C.DATASET.RELLIS_DIR in benchmarks/GSCNN-master/config.py The models are initialized by the weights pretrained on the ImageNet. You can download the pretrained models from here.

LiDAR Semantic Segmenation

SalsaNext

The SalsaNext is a fork from https://github.com/Halmstad-University/SalsaNext

To evaluate the dataset:

#!/bin/sh
export CUDA_VISIBLE_DEVICES="1"
cd /path/to/code/benchmarks/SalsaNext/train/tasks/semantic  
python infer2.py -d /path/to/RELLIS-3D -l /path/for/save/prediction -s test -m /path/to/pre-trained/model/folder

Dowload pre-trained model (Download 157MB)

To retrain the SalsaNext on RELLIS-3D:

export CUDA_VISIBLE_DEVICES="0,1"
cd /path/to/code/benchmarks/SalsaNext/train/tasks/semantic  
./train.py -d /path/to/RELLIS-3D  -ac ./config/arch/salsanext_ouster.yml -dc ./config/labels/rellis.yaml -n rellis -l ./logs -p ""

KPConv

The KPConv is a fork from https://github.com/HuguesTHOMAS/KPConv-PyTorch

To evaluate the dataset:

cd /path/to/code/benchmarks/KPConv-PyTorch-master
export PYTHONPATH=/path/to/code/benchmarks/KPConv-PyTorch-master/:$PYTHONPATH
python test_models.py

Configure benchmarks/KPConv-PyTorch-master/test_models.py before evaluation.

chosen_log = '/path/to/pretrained/model/folder'
config.sv_path = "/path/to/save/prediction"
config.data_path = "/path/to/RELLIS-3D"

Dowload pre-trained model (Download 1GB)

To retrain the KPConv on RELLIS-3D: To evaluate the dataset:

cd /path/to/code/benchmarks/KPConv-PyTorch-master
export PYTHONPATH=/path/to/code/benchmarks/KPConv-PyTorch-master/:$PYTHONPATH
python train_Rellis.py

Configure benchmarks/KPConv-PyTorch-master/train_Rellis.py before training.

data_path = "/path/to/RELLIS-3D"