This document summarizes the SNIPER method for efficient multi-scale object detection training. SNIPER generates image chips at multiple scales to handle scale variation, selects positive chips containing objects and hard negative chips lacking objects to train the model, and skips easy negative chips to speed up training. The method enables training on large batches with fewer computations than image pyramids. Experiments show SNIPER can train Faster R-CNN on MS COCO in 14 hours on a single GPU node while improving detection performance over single-scale training.