Revisiting Pixel-Level Contrastive Pre-Training on Scene Images