run_openmask3d_scannet200_eval.sh

#!/bin/bash
export OMP_NUM_THREADS=3  # speeds up MinkowskiEngine
set -e

# OPENMASK3D SCANNET200 EVALUATION SCRIPT
# This script performs the following in order to evaluate OpenMask3D predictions on the ScanNet200 validation set
# 1. Compute class agnostic masks and save them
# 2. Compute mask features for each mask and save them
# 3. Evaluate for closed-set 3D semantic instance segmentation

# --------
# NOTE: SET THESE PARAMETERS!
SCANS_PATH="/PATH/TO/SCANNET/SCANS"
SCANNET_PROCESSED_DIR="/PATH/TO/scannet_processed/scannet200"
# model ckpt paths
MASK_MODULE_CKPT_PATH="$(pwd)/resources/scannet200_val.ckpt"
SAM_CKPT_PATH="$(pwd)/resources/sam_vit_h_4b8939.pth"
# output directories to save masks and mask features
EXPERIMENT_NAME="scannet200"
OUTPUT_DIRECTORY="$(pwd)/output"
TIMESTAMP=$(date +"%Y-%m-%d-%H-%M-%S")
OUTPUT_FOLDER_DIRECTORY="${OUTPUT_DIRECTORY}/${TIMESTAMP}-${EXPERIMENT_NAME}"
MASK_SAVE_DIR="${OUTPUT_FOLDER_DIRECTORY}/masks"
MASK_FEATURE_SAVE_DIR="${OUTPUT_FOLDER_DIRECTORY}/mask_features"
SAVE_VISUALIZATIONS=false #if set to true, saves pyviz3d visualizations

# Paremeters below are AUTOMATICALLY set based on the parameters above:
SCANNET_LABEL_DB_PATH="${SCANNET_PROCESSED_DIR%/}/label_database.yaml"
SCANNET_INSTANCE_GT_DIR="${SCANNET_PROCESSED_DIR%/}/instance_gt/validation"
# gpu optimization
OPTIMIZE_GPU_USAGE=false

cd openmask3d

# 1.Compute class agnostic masks and save them
python class_agnostic_mask_computation/get_masks_scannet200.py \
general.experiment_name=${EXPERIMENT_NAME} \
general.project_name="scannet200" \
general.checkpoint=${MASK_MODULE_CKPT_PATH} \
general.train_mode=false \
model.num_queries=150 \
general.use_dbscan=true \
general.dbscan_eps=0.95 \
general.save_visualizations=${SAVE_VISUALIZATIONS} \
data.test_dataset.data_dir=${SCANNET_PROCESSED_DIR}  \
data.validation_dataset.data_dir=${SCANNET_PROCESSED_DIR} \
data.train_dataset.data_dir=${SCANNET_PROCESSED_DIR} \
data.train_dataset.label_db_filepath=${SCANNET_LABEL_DB_PATH} \
data.validation_dataset.label_db_filepath=${SCANNET_LABEL_DB_PATH} \
data.test_dataset.label_db_filepath=${SCANNET_LABEL_DB_PATH}  \
general.mask_save_dir=${MASK_SAVE_DIR} \
hydra.run.dir="${OUTPUT_FOLDER_DIRECTORY}/hydra_outputs/class_agnostic_mask_computation"
echo "[INFO] Mask computation done!"
# get the path of the saved masks
echo "[INFO] Masks saved to ${MASK_SAVE_DIR}."

# 2. Compute mask features
echo "[INFO] Computing mask features..."
python compute_features_scannet200.py \
data.scans_path=${SCANS_PATH} \
data.masks.masks_path=${MASK_SAVE_DIR} \
output.output_directory=${MASK_FEATURE_SAVE_DIR} \
output.experiment_name=${EXPERIMENT_NAME} \
external.sam_checkpoint=${SAM_CKPT_PATH} \
gpu.optimize_gpu_usage=${OPTIMIZE_GPU_USAGE} \
hydra.run.dir="${OUTPUT_FOLDER_DIRECTORY}/hydra_outputs/mask_features_computation"
echo "[INFO] Feature computation done!"

# 3. Evaluate for closed-set 3D semantic instance segmentation
python evaluation/run_eval_close_vocab_inst_seg.py \
--gt_dir=${SCANNET_INSTANCE_GT_DIR} \
--mask_pred_dir=${MASK_SAVE_DIR} \
--mask_features_dir=${MASK_FEATURE_SAVE_DIR} \