Methods¶

Six published methods, each ported stage by stage and checked against its reference code, the same weights loaded into both. For GSNet, the reference code is a third-party implementation (graspnet/graspness_unofficial). Five share the seed-based skeleton and its slots; HGGD reads RGB-D images and is a family of its own.

Seed-based skeleton · 5 methods¶

GraspNet-baseline · CVPR 2020 · depth

The baseline released with GraspNet-1Billion: PointNet++ seeds, 300 views, four depths.

equal to the reference ✓ authors' weights scored here
GSNet · ICCV 2021 · depth

Graspness-guided seeds on a sparse U-Net, one of 300 views, a grid of angles and depths.

equal to the reference ✓ trained here
Scale-Balanced-Grasp · CoRL 2022 · depth

Small grasps not drowned out: four crop radii and width-balanced losses.

equal to the reference ✓ authors' weights scored here
EconomicGrasp · ECCV 2024 · depth

One compact label per point and view; a head that classifies angle, depth and quality.

equal to the reference ✓ authors' weights scored here
FineGrasp · arXiv 2025 · depth

Surface normals as input, graspness per object, four crop radii fused by attention.

equal to the reference ✓ authors' weights scored here

Own family¶

HGGD · RA-L 2023 · RGB-D

A 2-D heat-map of grasp centres, completed in 3-D by a PointNet over local regions.

equal to the reference ✓ authors' weights scored here

What each method comes with¶

Computed from the configs and the metafiles.

Method	Released weights	Trained here	GraspNet-1Billion configs	GraspClutter6D configs
GraspNet-baseline	RealSense	—	RealSense	transfer from RealSense
GSNet	RealSense	RealSense, not published	RealSense	transfer from RealSense; training
Scale-Balanced-Grasp	RealSense	—	RealSense	transfer from RealSense
EconomicGrasp	Kinect, RealSense	—	Kinect, RealSense	transfer from Kinect, RealSense; training
FineGrasp	RealSense	—	RealSense	—
HGGD	RealSense, Kinect	—	Kinect, RealSense	transfer from RealSense

What each method ships in each slot

The component each method's config puts in every slot of its model; — where the method has no such slot.

Slot	GraspNet-baseline	GSNet	Scale-Balanced-Grasp	EconomicGrasp	FineGrasp	HGGD
Image encoder	—	—	—	—	—	built in
Scene encoder	`PointNet2`	`SparseUNet14`	`PointNet2Transformer`	`SparseUNet14`	`SparseUNet14`	—
Seed proposal	`SubsampledSeeds`	`GraspnessFPS`	`SubsampledSeeds`	`GraspnessFPS`	`GraspnessFPS`	—
View	`ObjectnessViewStage`	`TemplateViewStage`	`ObjectnessViewStage`	`TemplateViewStage`	`TemplateViewStage`	—
Assigner	`NearestLabelAssigner`	`NearestLabelAssigner`	`ScaleBalancedAssigner`	`CompactLabelAssigner`	`CompactLabelAssigner`	—
Grouper	`DepthCylinderGroup`	`CylinderGroup`	`DepthCylinderGroup`	`CylinderInteractionGroup`	`MultiScaleGroup`	—
Head bundle	`OperationHead`	`SWADHead`	`OperationHead`	`EconomicHead`	`EconomicHead`	—

Which one to run first¶

Every method page starts with a five-minute run from released weights.

GraspNet-baseline, Scale-Balanced-Grasp and HGGD need no extra install. HGGD reads a colour image registered to the depth image as well; the others read depth alone.

Method	Reads	Extra	Weights from	Licence	Demo
GraspNet-baseline	depth	none	Google Drive	non-commercial	10.5 s
GSNet	depth	`spconv`	Google Drive	non-commercial	14.8 s
Scale-Balanced-Grasp	depth	none	Google Drive	MIT, with non-commercial parts	10.5 s
EconomicGrasp	depth	`spconv`	GitHub release	MIT, with non-commercial parts	14.0 s
FineGrasp	depth	`spconv`	Hugging Face	Apache-2.0	14.8 s
HGGD	RGB-D	none	Tsinghua Cloud	MIT	14.5 s

Extra: spconv is installed as the extra that matches your PyTorch: spconv-cu118, spconv-cu121, spconv-cu124, spconv-cu126, spconv-cpu.
Licence: hover for the full text; each method page states it in full.
Demo is the whole grasppanda demo … --example command, measured on one NVIDIA RTX A6000 shared with other jobs; no run was timed on a CPU.
Every checkpoint was trained on GraspNet-1Billion, whose data and labels are under CC BY-NC-SA 4.0.