InstanceLoc

[arXiv 2021] Instance Localization for Self-supervised Detection Pretraining.

Main Results

Here we list the results on MSCOCO with the detector of R50-C4 and R50-FPN. Without any multi-crop/auto/random augmentation, our InsLoc outperforms many previous contrastive methods. In order to clearly clarify the improvements brought by InsLoc, the comparison to the corresponding baseline (i.e., MoCo-v2) is as follows:

Mask R-CNN R50-C4 1x:

Methods Epoch Box AP Mask AP Link
MoCo-v2 200 38.9 34.1 -
MoCo-v2 800 39.3 34.3 -
InsLoc 200 39.5 34.5 link
InsLoc 400 39.8 34.7 link

Mask R-CNN R50-C4 2x:

Methods Epoch Box AP Mask AP Link
MoCo-v2 200 40.7 35.6 -
MoCo-v2 800 41.2 35.8 -
InsLoc 200 41.4 35.9 link
InsLoc 400 41.8 36.3 link

Mask R-CNN R50-FPN 1x:

Methods Epoch Box AP Mask AP Link
MoCo-v2 200 39.8 36.1 -
MoCo-v2 800 40.4 36.4 -
InsLoc 200 41.4 37.1 link
InsLoc 400 42.0 37.6 link

Mask R-CNN R50-FPN 2x:

Methods Epoch Box AP Mask AP Link
MoCo-v2 200 41.7 37.6 -
MoCo-v2 800 42.5 38.2 -
InsLoc 200 43.2 38.7 link
InsLoc 400 43.3 38.8 link

More results are available in our paper.

Bibtex

@inproceedings{yang2021insloc,
  title={Instance Localization for Self-supervised Detection Pretraining},
  author={Yang, Ceyuan and Wu, Zhirong and Zhou, Bolei and Lin, Stephen},
  booktitle={arXiv preprint arXiv:2102.08318},
  year={2021},
}

GitHub