Junlin Chang

Point Ladder Tuning: Parameter-Efficient Hierarchical Adaptation for 3D Point Cloud Understanding

Junlin Chang, Longhao Zou, Rui Li

Fine-tuning pre-trained point-cloud backbones typically updates all parameters, resulting in substantial computation and memory overhead. More importantly, modern point backbones rely on aggressive tokenization and downsampling, which yields compact global tokens but irreversibly…Read more