# Fast and Accurate Object Detection on Asymmetrical Receptive Field

**Authors:** Liguo Zhou, Tianhao Lin, Alois Knoll  
**arXiv:** [2303.08995](https://arxiv.org/abs/2303.08995)  
**Date:** 2023-03-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.08995v1/

## Key points

- Object detection is important in industries like autonomous driving, robotics, and security
- Deep learning has significantly improved object detection accuracy and efficiency
- Challenges include highly accurate detection, multi-category object detection, real-time detection, and robustness to occlusions
- Proposed methods for improving mainstream object detection algorithms from the perspective of the evolution of one-stage and two-stage object detection algorithms
- Ways to enhance object detection accuracy by changing receptive fields are suggested
- The proposed model is based on YOLOv5 with modifications made to its head part through adding asymmetrical pooling layers
- Performance of the new model is compared with that of the original YOLOv5 model using several parameters and evaluated in four situations
- A confusion table for classification results helps understand true positive (TP), true negative (TN), false positive (FP) and false negative (FN) predictions made by the network.
- Spatial Pyramid Pooling technique enhances feature extraction by dividing an image into regions at multiple scales.
- Conclusion provides insights into improving mainstream object detection algorithms through modifying their structures and changing receptive fields resulting in enhanced accuracy without compromising speed.
- Future research directions include addressing challenges such as multi-category object detection and robustness to occlusions while ensuring real-time performance.

## Summary

Object detection is a crucial task in various industries such as autonomous driving, robotics, and security. With the advent of deep learning, object detection accuracy and efficiency have been significantly improved. To overcome challenges such as highly accurate detection, multi-category object detection, real-time detection and robustness to occlusions this paper proposes methods for improving mainstream object detection algorithms from the perspective of the evolution of one-stage and two-stage object detection algorithms. Additionally, this article suggests ways to enhance object detection accuracy by changing receptive fields. The proposed model is based on YOLOv5 (You Look Only Once), with modifications made to the structure of its head part through adding asymmetrical pooling layers. This modification improves algorithmic accuracy while maintaining speed. The performance of the new model is compared with that of the original YOLOv5 model using several parameters and evaluated in four situations. Furthermore, a confusion table for classification results helps understand true positive (TP), true negative (TN), false positive (FP) and false negative (FN) predictions made by the network. The article also introduces a technique called Spatial Pyramid Pooling that enhances feature extraction by dividing an image into regions at multiple scales. In conclusion, this paper provides insights into improving mainstream object detection algorithms through modifying their structures and changing receptive fields resulting in enhanced accuracy without compromising speed. Future research directions include addressing challenges such as multi-category object detection and robustness to occlusions while ensuring real-time performance.