search

Pointcept / PointTransformerV3

[CVPR'24 Oral] Official repository of Point Transformer V3 (PTv3)
MIT License
583 stars 30 forks source link

Feature extraction for point clouds #24

Open YunzeMan opened 2 months ago

YunzeMan commented 2 months ago

Hi, thanks for your great work. I am trying to use your PTv3 as a frozen feature extractor for new 3D point clouds, like how people use CLIP/DINOv2 for images.

Could you share some simple scripts or ideas about how we can achieve this with your codebase? This can be very helpful.

Gofinge commented 2 months ago

Hi, I think you can refer to our framework classes like MaskedSceneContrast (https://github.com/Pointcept/Pointcept/blob/main/pointcept/models/masked_scene_contrast/masked_scene_contrast_v1m1_base.py#L25) or DefaultSegmentorV2 (https://github.com/Pointcept/Pointcept/blob/main/pointcept/models/default.py#L35).

These framework model classes contain an element defined as backbone and can be initiated by backbone config (refer to our configs). Then you can freeze the parameter of the backbone in a framework model class and include pre-training context tasks.