ObjectNet is the same size as the ImageNet test set (50,000 images), and by design does not comepaired withatraining setinordertoencourage generalization.

However,thecompression only in the spatial domain suffers from a dramatic performance drop without finetuning and is not robust to noise, as the noise in the spatial domain can easily confuse the pruning criteria, leading to some parameters/channels being pruned incorrectly.

We thank all the reviewers for the valuable comments and suggestions. Besides, we indeed use dropout as in NoisyStudent (the paper you mentioned) to help generalization. We also combine SemiNAS with other NAS algorithm (e.g., Regularized Evolution) and We will add such experiments in the new version. SemiNAS (RE) consuming 2000 pairs to compare with RE under the same number of queries, and it achieves 94.03% CIFAR-10, there exist some differences. It runs each model for 3 times and collect the 3 results to reduce the variance.

Wefirst implement global augmentation approaches including random rotation and random scaling on two LiDAR scans separately and thenconcatenate themfortraining. The more copies the better segmentation performance as shown in ' 1, 2, 3' in the table, which indicates the effectiveness of the approach in enriching data distribution. In this section, we conducted experiments to analyze how PolarMix benefits LiDAR point cloud learning. As a comparison, PolarMix is more robust to the instance spatial location without much performance drop. PolarMix improves the robustness of the baseline clearly with respect to the angular variations of instances (i.e.