32nd Conference on Neural Information Processing Systems (NeurIPS 2018)

By Robert J. Wang et al.

Read the original document by opening this link in a new tab.

Table of Contents

1. Abstract
2. Introduction
3. Two-Way Dense Layer
4. Stem Block
5. Dynamic Number of Channels in Bottleneck Layer
6. Transition Layer without Compression
7. Composite Function
8. PeleeNet: An Efficient Feature Extraction Network
9. Ablation Study
10. Results on ImageNet ILSVRC 2012
11. Speed on Real Devices
12. Pelee: A Real-Time Object Detection System
13. Overview

Summary

The document discusses the development of Pelee, a real-time object detection system that combines PeleeNet with Single Shot MultiBox Detector (SSD) method. It introduces the architecture of PeleeNet, focusing on its design choices and performance evaluation on Stanford Dogs dataset and ImageNet ILSVRC 2012. The document also evaluates the speed of PeleeNet on real devices such as NVIDIA TX2 and iPhone 8. Furthermore, it presents the optimization of Pelee for SSD, achieving high mAP on PASCAL VOC2007 and MS COCO dataset, outperforming YOLOv2 in terms of accuracy and model size.