search

Pang-Yatian / Point-MAE

[ECCV2022] Masked Autoencoders for Point Cloud Self-supervised Learning
MIT License
448 stars 55 forks source link

ModelNet(8k) accuracy of 1024 PCs #45

Closed Harr7y closed 6 months ago

Harr7y commented 7 months ago

Hi, I test the accuracy of the finetuning ModelNet40(8k) checkpoint on 1024 points, and the result shows that the accuracy is about 68.52% without voting. I am curious if the model's after the finetuning with 8k PCs is only suitable for the case of 8k PC input?

Model: ModelNet40 (8k) Input : 1024 PCs -> Accuracy: 68.52% Input : 8192 PCs -> Accuracy: 93.40%

The number of points is changed through N_POINTS in ModelNet40.yaml

Or is this something wrong with my command: command: python main.py --config cfgs/finetune_modelnet_8k.yaml --test --exp_name test --ckpts /Point-MAE/checkpoint/modelnet_8k.pth

Pang-Yatian commented 6 months ago

Yes, it is normal. We also tested this finetune pipeline before and found similar results. I guess the point cloud density is different from the image resolution, so this way cannot work.

Harr7y commented 6 months ago

Thanks for your reply. :)