Feature representation via self-supervised learning has reached remarkable\nsuccess in image-level contrastive learning, which brings impressive\nperformances on image classification tasks. While image-level feature\nrepresentation mainly focuses on contrastive learning in single instance, it\nignores the objective differences between pretext and downstream prediction\ntasks such as object detection and instance segmentation. In order to fully\nunleash the power of feature representation on downstream prediction tasks, we\npropose a new end-to-end self-supervised framework called InsCon, which is\ndevoted to capturing multi-instance information and extracting cell-instance\nfeatures for object recognition and localization. On the one hand, InsCon\nbuilds a targeted learning paradigm that applies multi-instance images as\ninput, aligning the learned feature between corresponding instance views, which\nmakes it more appropriate for multi-instance recognition tasks. On the other\nhand, InsCon introduces the pull and push of cell-instance, which utilizes cell\nconsistency to enhance fine-grained feature representation for precise boundary\nlocalization. As a result, InsCon learns multi-instance consistency on semantic\nfeature representation and cell-instance consistency on spatial feature\nrepresentation. Experiments demonstrate the method we proposed surpasses MoCo\nv2 by 1.1% AP^{bb} on COCO object detection and 1.0% AP^{mk} on COCO instance\nsegmentation using Mask R-CNN R50-FPN network structure with 90k iterations,\n2.1% APbb on PASCAL VOC objection detection using Faster R-CNN R50-C4 network\nstructure with 24k iterations.\n