chenhongyiyang/ccop

Contrastive Object-level Pre-training with Spatial Noise Curriculum Learning

ccop's Issues

One missing configuration setting

Thanks for your great work and releasing the code!

During reproducing the experimental results, I noticed a tiny discrepancy between your code and the standard Mask R-CNN fine-tuning:

In your configuration file benchmarks/detection/configs/coco_R_50_FPN_1x_infomin.yaml, the configuration MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION is not set. As a result, Detectron2 will use the default setting MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION = 14, from https://github.com/facebookresearch/detectron2/blob/main/detectron2/config/defaults.py. However, this does not match the standard Mask R-CNN setting MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION = 7. Please see https://github.com/facebookresearch/detectron2/blob/main/configs/Base-RCNN-FPN.yaml.

I'm not sure which Detectron2 version you were previously using, so this difference could be a result of Detectron2 updates as well.

selective_search.py is empty

hi, @ChenhongyiYang , why the mentioned file is empty, does that affect the self-supervised pretraining?