Abstract:
Arguably, one of the deep learning’s greatest achievements is transfer learning. It
is proven that if we use a rich source dataset to pre-train a network (ImageNet in
2D) and after we use a smaller target dataset to be fine-tuned, it can help to boost
performance. It has been used in many applications including language and
vision. In 3D scene understanding, there have been few works using this method.
Because annotating 3D data is difficult. In this work, we aim to further facilitate
research on 3D representation learning. To achieve this goal, we select different
datasets and downstream tasks to find the effectiveness of unsupervised pre training on a large and rich source dataset of a 3D scene. The results we obtain
are encouraging. we are using a unified backbone, source dataset, and contrastive
objective for unsupervised pre-training, and further supervised downstream tasks
are performed. Our method is achieving the almost same result in half time of
recent works in both pre-training and downstream tasks like semantic
segmentation, and instance segmentation.
