Skip to content

Latest commit

 

History

History
93 lines (77 loc) · 4.32 KB

S3DIS.md

File metadata and controls

93 lines (77 loc) · 4.32 KB
Raw

S3DIS

Download S3DIS Dataset and extract Stanford3dDataset_v1.2_Aligned_Version.zip to $S3DIS_DIR/data, where $S3DIS_DIR is set to dataset directory.

To fix some issues with the dataset as reported in issue #29, apply path S3DIS_fix.diff with:

cp S3DIS_fix.diff $S3DIS_DIR/data; cd $S3DIS_DIR/data; git apply S3DIS_fix.diff; rm S3DIS_fix.diff; cd -

Define $S3DIS_DIR as the location of the folder containing /data

SPG with Handcrafted Partition

To compute the partition with handcrafted features run:

python partition/partition.py --dataset s3dis --ROOT_PATH $S3DIS_DIR --voxel_width 0.03 --reg_strength 0.03

Then, reorganize point clouds into superpoints by:

python learning/s3dis_dataset.py --S3DIS_PATH $S3DIS_DIR

To train from scratch on the all 6 folds on the handcrafted partition, run:

for FOLD in 1 2 3 4 5 6; do \
CUDA_VISIBLE_DEVICES=0 python learning/main.py --dataset s3dis --S3DIS_PATH $S3DIS_DIR --cvfold $FOLD --epochs 350 \
--lr_steps '[275,320]' --test_nth_epoch 50 --model_config 'gru_10_0,f_13' --ptn_nfeat_stn 14 --nworkers 2 \
--pc_attribs xyzrgbelpsvXYZ --odir "results/s3dis/best/cv${FOLD}" --nworkers 4; \
done

Our trained networks can be downloaded here. Unzip the folder (but not the model.pth.tar themselves) and place them in the code folder results/s3dis/pretrained/.

To test these networks on the full test set, run:

for FOLD in 1 2 3 4 5 6; do \
CUDA_VISIBLE_DEVICES=0 python learning/main.py --dataset s3dis --S3DIS_PATH $S3DIS_DIR --cvfold $FOLD --epochs -1 --lr_steps '[275,320]' \
--test_nth_epoch 50 --model_config 'gru_10_0,f_13' --ptn_nfeat_stn 14 --nworkers 2 --pc_attribs xyzrgbelpsvXYZ --odir "results/s3dis/pretrained/cv${FOLD}" --resume RESUME; \
done

SSP+SPG: SPG with learned partition

To learn the partition from scratch run:

python supervized_partition/graph_processing.py --ROOT_PATH $S3DIS_DIR --dataset s3dis --voxel_width 0.03; \

for FOLD in 1 2 3 4 5 6; do \
    python ./supervized_partition/supervized_partition.py --ROOT_PATH $S3DIS_DIR  --cvfold $FOLD \
    --odir results_partition/s3dis/best --epochs 50 --reg_strength 0.1 --spatial_emb 0.2  \
    --global_feat eXYrgb --CP_cutoff 25; \
done

Or download our trained weights here in the folder results_partition/s3dis/pretrained, unzipped and run the following code:

for FOLD in 1 2 3 4 5 6; do \
    python ./supervized_partition/supervized_partition.py --ROOT_PATH $S3DIS_DIR  --cvfold $FOLD --epochs -1 \
    --odir results_partition/s3dis/pretrained --reg_strength 0.1 --spatial_emb 0.2 --global_feat eXYrgb \
    --CP_cutoff 25 --resume RESUME; \
done

To evaluate the quality of the partition, run:

python supervized_partition/evaluate_partition.py --dataset s3dis --folder pretrained --cvfold 123456

Then, reorganize point clouds into superpoints with:

python learning/s3dis_dataset.py --S3DIS_PATH $S3DIS_DIR --supervized_partition 1  -plane_model_elevation 1

Then to learn the SPG models from scratch, run:

for FOLD in 1 2 3 4 5 6; do \
	CUDA_VISIBLE_DEVICES=0 python ./learning/main.py --dataset s3dis --S3DIS_PATH $S3DIS_DIR --batch_size 5 \
  --cvfold $FOLD --epochs 250 --lr_steps '[150,200]' --model_config "gru_10_0,f_13" --ptn_nfeat_stn 10 \
  --nworkers 2 --spg_augm_order 5 --pc_attribs xyzXYZrgbe --spg_augm_hardcutoff 768 --ptn_minpts 50 \
  --use_val_set 1 --odir results/s3dis/best/cv$FOLD; \
    done;

Or use our trained weights with --epochs -1 and --resume RESUME:

for FOLD in 1 2 3 4 5 6; do \
	CUDA_VISIBLE_DEVICES=0 python ./learning/main.py --dataset s3dis --S3DIS_PATH $S3DIS_DIR --batch_size 5 \
  --cvfold $FOLD --epochs -1 --lr_steps '[150,200]' --model_config "gru_10_0,f_13" --ptn_nfeat_stn 10 \
  --nworkers 2 --spg_augm_order 5 --pc_attribs xyzXYZrgbe --spg_augm_hardcutoff 768 --ptn_minpts 50 \
  --use_val_set 1 --odir results/s3dis/pretrained_SSP/cv$FOLD --resume RESUME; \
    done;

Note that these weights are specifically adapted to the pretrained model for the learned partition. Any change to the partition might decrease their performance.