# USB: Universal-Scale Object Detection Benchmark

**Authors:** Yosuke Shinya  
**arXiv:** [2103.14027](https://arxiv.org/abs/2103.14027)  
**Date:** 2021-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2103.14027v1/

## Key points

- Object detection benchmarks like COCO are important for evaluating algorithm performance
- Existing benchmarks have limitations in scale variation and fair comparison protocols
- The Universal-Scale object detection Benchmark (USB) addresses these issues
- USB incorporates COCO with Waymo Open Dataset and Manga109-s dataset for comprehensive evaluation
- USB protocols define thresholds for training epochs and evaluation image resolutions to ensure fair comparison
- UniverseNets, fast and accurate object detectors, were designed using the proposed benchmark
- UniverseNets outperformed baselines on USB and achieved state-of-the-art results on existing benchmarks
- UniverseNets achieved 54.1% AP on COCO test-dev dataset with 20 epochs of training
- UniverseNets secured top result in Waymo Open Dataset Challenge 2020 2D detection category and NightOwls Detection Challenge 2020 all objects track
- The paper presents a novel benchmarking framework that addresses shortcomings in existing benchmarks by incorporating additional datasets with variations in scale and image domains
- USB protocols enable fair comparison between different methods
- Code for UniverseNets is publicly available at https://github.com/shinya7y/UniverseNet

## Summary

In the field of object detection, benchmarks such as COCO are essential for evaluating the performance of different algorithms. However, existing benchmarks have limitations in terms of scale variation and protocols for fair comparison. To address these issues, this paper introduces the Universal-Scale object detection Benchmark (USB). The USB benchmark incorporates COCO with two additional datasets: the Waymo Open Dataset and Manga109-s dataset. This integration allows for variations in object scales and image domains, providing a more comprehensive evaluation framework. To ensure fair comparison among different methods, the authors propose USB protocols that define multiple thresholds for training epochs and evaluation image resolutions. These protocols enable researchers to evaluate their models under various conditions and compare them accurately. Using the proposed benchmark, the authors designed fast and accurate object detectors called UniverseNets. These detectors outperformed all baselines on USB and achieved state-of-the-art results on existing benchmarks. Specifically, UniverseNets achieved an impressive 54.1% Average Precision (AP) on the COCO test-dev dataset with 20 epochs of training. They also achieved the top result among single-stage detectors in the Waymo Open Dataset Challenge 2020 2D detection category and secured first place in the NightOwls Detection Challenge 2020 all objects track. Overall, this paper presents a novel benchmarking framework that addresses shortcomings in existing benchmarks by incorporating additional datasets with variations in scale and image domains. The proposed USB protocols enable fair comparison between different methods, leading to the development of fast and accurate object detectors like UniverseNets that achieve state-of-the-art results across multiple benchmarks. The code for UniverseNets is publicly available at https://github.com/shinya7y/UniverseNet .