Beyond Context Bias: Adaptive Instance Placement for Robust 3D Instance Segmentation

Paper · Project Page · BibTeX

This repository contains the official implementation of Adaptive Instance Placement for Robust 3D Instance Segmentation (ACGP).

Overall framework of ACGP. Given a 3D scene and an instance database, ACGP selects instances via category-balanced sampling, queries their category-conditioned placement types, and places them at geometrically valid locations using occupancy- and support-aware validation to generate augmented scenes. The pseudo-code on the right summarizes the complete pipeline of instance-level augmented scene generation.

The core ACGP instance placement implementation can be found in pointcept/datasets/instance_augmentor_occupancy.py.

📢 News

Update: ACGP achieves the SOTA performance on ScanNet200 test set for 3D instance segmentation benchmark. Test scores accessed on 12 July, 2026. ScanNet200 test set (The results are provided by official based on this repo)

ACGP achieves the SOTA performance on ScanNet++ V2 test set for 3D instance segmentation. Test scores accessed on 10 July, 2026. ScanNet++ V2 test set

💾 ACGP Trained Results

Benchmark	Model	mAP	AP50	AP25	Tensorboard	Exp Record	Model
ScanNet++ Val	Volt-SPFormer	37.2	56.3	64.6	Link	Link	Link
	SGIFormer	35.4	52.4	61.4	Link	Link	Link
ScanNet200 Val	Volt-SPFormer	39.6	50.2	55.2	Link	Link	Link
	SGIFormer-L	32.9	43.3	49.3	Link	Link	Link
	ISBNet	26.2	34.4	39.2	Link	Link	Link
	TD3D	24.2	36.2	42.2	-	-	Link
ScanNet Val	SGIFormer-L	61.6	81.6	89.2	Link	Link	Link
	ISBNet	56.1	73.6	81.8	Link	Link	Link
	SoftGroup	44.0	65.3	77.9	Link	Link	Link
	TD3D	47.6	71.3	82.1	-	-	Link
S3DIS Area5	ISBNet	58.2	71.3	77.7	Link	Link	Link
	SoftGroup	53.5	66.7	73.7	Link	Link	Link
	TD3D	53.1	68.2	75.9	-	-	Link
STPLS3D Val	ISBNet	52.1	68.9	76.0	Link	Link	Link

Setup

This repository is built on top of Pointcept and incorporates components from SGIFormer and Volt for instance segmentation.

Dependencies

We recommend using uv, a fast Python package and environment manager, to install the environment.

To install uv on macOS and Linux, run:

curl -LsSf https://astral.sh/uv/install.sh | sh

Then set up the environment with:

# Make sure to load CUDA 12.6 beforehand
# This will automatically create a virtual environment (.venv) and install dependencies from pyproject.toml
uv sync
source .venv/bin/activate

Data Preprocessing

Follow the dataset setup instructions in the Pointcept README.

Indoor Datasets

Preprocessing for indoor datasets is identical to Pointcept.

Outdoor Datasets

Preprocessing for outdoor datasets is identical to ISBNet.

Instance Segmentation

First, run the preprocessing script to generate superpoints for ScanNet and ScanNet200.

python pointcept/datasets/preprocessing/scannet/preprocess_superpoints.py --dataset_root ${RAW_SCANNET_DIR} --output_root ${PROCESSED_SCANNET_DIR}

Download the pretrained Volt-S backbone weights from HuggingFace

mkdir -p weights
curl -L -o weights/volt-small-scannet.pth https://huggingface.co/KadirYilmaz/Volt/resolve/main/Volt_experiments/joint_training_small/scannet/model/model_last.pth
curl -L -o weights/volt-small-scannet200.pth https://huggingface.co/KadirYilmaz/Volt/resolve/main/Volt_experiments/joint_training_small/scannet200/model/model_last.pth

Alternatively you can train them yourself using the corresponding configs above.

Then, run the training script with the insseg-spformer-volt-S-0-base config for scannet/scannet200

### ScanNet
sh scripts/train.sh -g 4 -d scannet -c insseg-spformer-volt-S-0-base -n insseg-volt
### ScanNet200
sh scripts/train.sh -g 4 -d scannet200 -c insseg-spformer-volt-S-0-base -n insseg-volt

Citation

If you use our work in your research, please use the following BibTeX entry.

@misc{yang2026acgp,
  title  = {Beyond Context Bias: Adaptive Instance Placement for Robust 3D Instance Segmentation},
  author = {Rongkun Yang, Ye Zhang, Longguang Wang, Zhiheng Fu, Lian Xu, Yulan Guo},
  year   = {2026},
  url    = {https://github.com/SYSU-SAIL/ACGP}
}

Acknowledgements

Code is built based on Volt, PointCept, and SGIFormer. We sincerely thank the authors for sharing their code.