SSD Paper Reading Record

SSD: Single Shot MultiBox Detector

Paper: SSD: Single Shot MultiBox Detector

Pass 1

SSD is a multiple categories detector. The model uses multi-scale convolutional bounding box attached with multiple feature maps at the top of the network. It enables models to efficiently track the possible shapes of target object. SSD significantly outperforms other state-of-the-art object detector models, such as Faster R-CNN, while being 3 times faster. As a monolitic and relatively simple object detection model, a promising future of SSD is its application as a part of an RNN structure to simultaneously detect and track objects in videos.

Pass 2

The state-of-the-art inspection systems before the advent of SSD conducted object detection in the following three stages: bounding box prediction, pixel and feature resampling, and classification. Though getting satisfactory classification results, these detections systems required too much computational power and performed slowly for real-time applications. SSD, unlike those previous systems, significantly reduced computaional demands by discarding the resampling step. SSD used a small convolutional filter for object prediction and bounding box offset. These design features made end-to-end training possible. The research team compared SSD with other state-of-the-art models on several different datasets such as PASCAL VOC, COCO, and ILSVRC.