In bash files, there is --config option for choosing the Open3DIS configuration
1) Extract 2D masks and first stage feature from RGB-D sequences:
sh scripts/grounding_2d.sh <config>
2) Generate 3D instances from 2D masks:
sh scripts/generate_3d_inst.sh <config>
3) Refine second stage feature from 3D instances:
sh scripts/refine_grounding_feat.sh <config>
After refine grounded features, re-run step 2 to finalize the 3D output masks
4) Run interactive visualization (required Pyviz3D):
sh scripts/vis.sh
5) Run evaluation:
We provided custom evaluation script
sh scripts/eval.sh
6) Misc:
Promptable Segmentation
sh scripts/text_query.sh <config> <text_query>
Maskwise feature computation memory efficient
sh scripts/maskwise_vocab.sh
Class-agnostic evaluation
sh scripts/eval_classagnostic.sh
Reproduce OpenSUN3D challenge 3D instance retrieval
# Generating results
sh scripts/opensun3d.sh
# Submit results
python tools/submit_opensun3d.py
Submit ScanNet++ 3D instance segmentation benchmark
# Step 1-4 with 'configs/scannetpp_benchmark_instance_test.yaml'
# Submit results
sh scripts/submit_scannetpp_benchmark.sh configs/scannetpp_benchmark_instance_test.yaml
OpenYOLO3D achieves remarkable results, we provide a reproducible version of OpenYOLO3D based on Open3DIS codebase
sh scripts/reproduce_openyolo3d.sh <config>