Unsupervised Object-Level Representation Learning from Scene Images Supplementary Material

The results are averaged across five independent runs. The learning rate is decayed by 0.2 at 12 and 16 epochs. The learning rate is initialized as 0.02 with a linear warmup for The implementation details of our most essential image-level baseline, i.e., BYOL [5], are provided Our reproduced results vs. existing results for BYOL. All are based on 800-epoch pre-training on COCO with ResNet-50. Figure 2 visualizes more attention maps generated by BYOL and ORL.