Tech Report: One-stage Lightweight Object Detectors

TL;DR

This paper introduces new one-stage lightweight object detectors optimized for both GPU and CPU, achieving a better balance of accuracy and latency on the MS COCO dataset.

Contribution

It proposes novel backbone and feature pyramid network architectures that improve speed and accuracy trade-offs compared to existing models like YOLOX-tiny.

Findings

Proposed GPU-target backbone outperforms YOLOX-tiny by 1.43x in speed.

Achieves 0.5 mAP improvement on MS COCO dataset.

Demonstrates effective trade-offs between parameters, Gflops, latency, and accuracy.

Abstract

This work is for designing one-stage lightweight detectors which perform well in terms of mAP and latency. With baseline models each of which targets on GPU and CPU respectively, various operations are applied instead of the main operations in backbone networks of baseline models. In addition to experiments about backbone networks and operations, several feature pyramid network (FPN) architectures are investigated. Benchmarks and proposed detectors are analyzed in terms of the number of parameters, Gflops, GPU latency, CPU latency and mAP, on MS COCO dataset which is a benchmark dataset in object detection. This work propose similar or better network architectures considering the trade-off between accuracy and latency. For example, our proposed GPU-target backbone network outperforms that of YOLOX-tiny which is selected as the benchmark by 1.43x in speed and 0.5 mAP in accuracy on NVIDIA GeForce RTX 2080 Ti GPU.