Closed bigKoki closed 10 months ago
Hi, thanks for your interest in this work. Actually, the great performance of multi-view image-based learning models relies on the quality of the rendered images. If the image is rendered from mesh models, the teacher shows much better performances on tasks like classification/retrieval. If there is no available mesh models and thus the image can only be rendered from raw points, then the teacher's performance is worsen than point-based networks. However, as validated in our work, in this setting, even if the teacher is weak, the distillation process can still boost the student point-based networks.
您好,感谢您对这项工作的关注。实际上,基于多视图图像的学习模型的出色性能取决于渲染图像的质量。如果图像是从网格模型渲染的,则教师在分类/检索等任务上表现出更好的表现。如果没有可用的网格模型,因此只能从原始点渲染图像,那么教师的性能会比基于点的网络更差。然而,正如我们的工作所验证的那样,在这种情况下,即使教师很弱,蒸馏过程仍然可以促进学生基于点的网络。
sorry for the late reply,thanks for your explain!!!
firstly,your work is very creative!!! But i have some question.In 3D tasks,multi-view methods are better than other methods.In your work,the teacher network boost the student network,if i only use trained teacher network for testing,will the result be better than student network? thanks!!!!