# 2nd Place Solution for Waymo Open Dataset Challenge -- Real-time 2D Object Detection

**Authors:** Yueming Zhang, Xiaolin Song, Bing Bai, Tengfei Xing, Chao Liu, Xin Gao, Zhihui Wang, Yawei Wen, Haojin Liao, Guoshan Zhang, Pengfei Xu  
**arXiv:** [2106.08713](https://arxiv.org/abs/2106.08713)  
**Date:** 2021-06-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.08713v1/

## Key points

- Accurately detecting and recognizing vehicles, pedestrians, and cyclists from images is crucial in autonomous driving
- Real-time predictions pose challenges for convolutional network models
- A real-time method for detecting 2D objects from images is introduced
- Approach involves aggregating popular one-stage object detectors and training models with various input strategies independently
- Better performance in accurately detecting objects of different sizes across multiple scales, particularly small objects
- TensorRT is leveraged for model acceleration to optimize inference time
- Achieved impressive results in the Waymo Open Dataset Challenges, ranking second place on the leaderboard with a L1 mAP score of 75.00% and a L2 mAP score of 69.72%
- Efficient processing capabilities with a latency of 45.8ms per frame when running on an Nvidia Tesla V100 GPU
- Comprehensive solution for real-time object detection in autonomous driving systems that combines multiple object detectors and optimizes inference time using TensorRT
- Contributes to advancements in autonomous driving technology and paves the way for safer and more efficient self-driving vehicles

## Summary

In the field of autonomous driving, accurately detecting and recognizing vehicles, pedestrians, and cyclists from images is crucial. However, achieving high accuracy predictions in real-time poses new challenges for convolutional network models. In this report, a real-time method for detecting 2D objects from images is introduced. The approach involves aggregating several popular one-stage object detectors and training models with various input strategies independently. This allows for better performance in accurately detecting objects of different sizes across multiple scales, particularly small objects. To optimize the inference time of the detection pipeline, TensorRT is leveraged for model acceleration. The proposed detection framework was evaluated in the real-time 2D detection track of the Waymo Open Dataset Challenges. It achieved impressive results, ranking second place on the leaderboard with a L1 mean average precision (mAP) score of 75.00% and a L2 mAP score of 69.72%. Additionally, the framework demonstrated efficient processing capabilities with a latency of 45.8ms per frame when running on an Nvidia Tesla V100 GPU. Overall, this report presents a comprehensive solution for real-time object detection in autonomous driving systems that combines multiple object detectors and optimizes inference time using TensorRT to achieve accurate multi-scale detection even for small objects. These findings contribute to advancements in autonomous driving technology and pave the way for safer and more efficient self-driving vehicles.