# Tiny-YOLO object detection supplemented with geometrical data

**Authors:** Ivan Khokhlov, Egor Davydenko, Ilia Osokin, Ilya Ryakin, Azer Babaev, Vladimir Litvinenko, Roman Gorbachev  
**arXiv:** [2008.02170](https://arxiv.org/abs/2008.02170)  
**Date:** 2020-08-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2008.02170v1/

## Key points

- Authors propose a method to improve object detection precision using prior knowledge about scene geometry
- Focus specifically on autonomous robots and assume the scene is a plane with objects placed on it
- Modify YOLOv3-tiny by introducing a scale channel (S) based on robot's dimensions and camera inclination angles
- Experiments show that this approach outperforms standard RGB-based detection methods with minimal computational overhead
- Potential applications in various fields involving autonomous robots and object detection tasks
- Findings presented in a 5-page paper accompanied by 5 figures published in the 2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring)
- Demonstrates how incorporating prior knowledge about scene geometry improves object detection accuracy while minimizing computational costs.

## Summary

In their paper titled "Tiny-YOLO object detection supplemented with geometrical data," authors Ivan Khokhlov, Egor Davydenko, Ilia Osokin, Ilya Ryakin, Azer Babaev, Vladimir Litvinenko, and Roman Gorbachev propose a method to improve the detection precision (mAP) in object detection using prior knowledge about the scene geometry. They assume that the scene is a plane with objects placed on it and focus specifically on autonomous robots. To predict the spatial scale for each pixel of the input frame, the authors take into account the robot's dimensions and the inclination angles of the camera. With this information, they modify YOLOv3-tiny and introduce a scale channel (referred to as S) to enhance detection performance. Their experiments show that this approach outperforms standard RGB-based detection methods while incurring only a small computational overhead. The proposed method has potential applications in various fields involving autonomous robots and object detection tasks. The authors present their findings in a 5-page paper accompanied by 5 figures which was published in the 2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring). This paper demonstrates how incorporating prior knowledge about scene geometry can improve object detection accuracy while minimizing computational costs.