Model checkpoints

We provide here results and checkpoints for various models trained on the 2D and 3D modalities of 3DCoMPaT++. We provide models and results for:

• Coarse and fine-grained part

• Material segmentation (coarse-grained)

• Shape classification

🦁 2D Models

We provide here some 2D models trained on the 3DCoMPaT dataset. For all models, you must include the loader files from the repo_root/loaders/2D in the active folder (you can simply use a symlink).

Results.

A. Fine-grained part segmentation.

Model	MIOU	mPrecision	ckpt
SegFormer	52.24	97.01	gdrive

B. Coarse-grained part segmentation.

Model	MIOU	mPrecision	ckpt
SegFormer	73.35	99.09	gdrive

C. Material segmentation.

Model	MIOU	mPrecision	ckpt
SegFormer	82.45	95.82	gdrive

Classification.

Point-cloud classification.

Model	Accuracy top-1	Accuracy top-1	ckpt
resnet18	76.269	97.06	gdrive
resnet50	90.203	99.09	gdrive

🐯 3D Models

We provide here some 3D models trained on the 3DCoMPaT dataset. This repo includes the code for 3d Shape Classification and Part Segmentation on 3DCoMPaT dataset for both coarse and fine grained versions using prevalent 3D vision algorithms, including PointNet++, DGCNN, PCT, PointStack, and CurveNet in pytorch.

You can find the pretrained models and log files in gdrive.

Results.

A. Fine-grained part segmentation.

Model	Number of points	Accuracy	Shape-aware mIOU	Shape-agnostic mIOU	ckpt
PCT	2048	70.49	81.31	49.09	gdrive
PointNet2 partseg_ssg	2048	71.09	80.01	50.39	gdrive
Curvenet	2048	72.49	81.37	53.09	gdrive

B. Coarse-grained part segmentation.

Model	Number of points	Accuracy	Shape-aware mIOU	Shape-agnostic mIOU	ckpt
PCT	2048	80.64	75.48	66.95	gdrive
PointNet2 partseg_ssg	2048	84.72	77.98	73.79	gdrive
Curvenet	2048	86.01	80.64	76.32	gdrive

C. Shape classification.

Model	Number of points	Accuracy	ckpt
DGCNN	2048	78.85	gdrive
PCT	2048	68.88	gdrive
PointNet2 cls_msg	2048	84.10	gdrive
PointStack	2048	83.04	gdrive
Curvenet	2048	85.14	gdrive
PointNext	2048	83.01	gdrive