We explore how pre-training a model to infer depth from a single image compares to pre-training the model for a semantic task, e.g. ImageNet classification, for the purpose of downstream transfer to semantic segmentation. The question of whether pre-training on geometric tasks is viable for downstream transfer to semantic tasks is important for two reasons, one practical and the other scientific. In practice, if it were viable, one could reduce pre-training costs and bias due to human annotation at scale. If, however, it were not, then that would affirm human annotation as an inductive vehicle so powerful to justify the annotation effort. Yet the bootstrapping question would still be unanswered: How did the ability to assign labels to semantically coherent regions emerge? If pre-training on a geometric task was sufficient to prime a notion of 'object', leveraging the regularities of the environment (what Gibson called 'detached objects'), that would reduce the gap to semantic inference as a matter of aligning labels, which could be done with few examples. To test these hypotheses, we have designed multiple controlled experiments that require minimal fine-tuning, using common benchmarks such as KITTI, Cityscapes, and NYU-V2: We explore different forms of supervision for depth estimation, training pipelines, and data resolutions for semantic fine-tuning. We find that depth pre-training exceeds performance relative to ImageNet pre-training on average by 5.8% mIoU and 5.2% pixel accuracy. Surprisingly, we find that optical flow estimation, which is a closely related task to depth estimation as it optimizes the same photometric reprojection error, is considerably less effective.

翻译：暂无翻译