# Object Detection in Equirectangular Panorama

**Authors:** Wenyan Yang, Yanlin Qian, Francesco Cricri, Lixin Fan, Joni-Kristian Kamarainen  
**arXiv:** [1805.08009](https://arxiv.org/abs/1805.08009)  
**Date:** 2018-05-21  
**Summary page:** https://www.summarizepaper.com/en/paper/1805.08009v1/

## Key points

- The paper introduces a novel approach for object detection in equirectangular panoramas using a multi-projection variant of the YOLO detector.
- The authors address challenges such as lack of annotated training data, high-resolution imagery, and severe geometric distortions near the panorama projection poles.
- They leverage training examples from conventional datasets like ImageNet and COCO to overcome the lack of annotated training data specific to equirectangular panoramas.
- They employ low-resolution images that require moderate GPU computing power and memory, making it feasible to process high-resolution equirectangular panoramas efficiently.
- The key innovation lies in the multi-projection YOLO detector which handles projection distortions by creating multiple stereographic sub-projections for more accurate object detection.
- In experiments, their multi-projection YOLO outperforms other state-of-the-art detectors like Faster RCNN in terms of accuracy for object detection in equirectangular panoramas.
- The multi-projection YOLO achieves the best accuracy when using low resolution input.
- Overall, the paper presents a comprehensive solution for object detection in equirectangular panoramas by addressing challenges related to training data availability, computational efficiency, and handling geometric distortions.

## Summary

The paper titled "Object Detection in Equirectangular Panorama" introduces a novel approach for object detection in high-resolution equirectangular panoramas using a multi-projection variant of the YOLO detector. To address the challenges associated with object detection in equirectangular panoramas such as the lack of annotated training data, high-resolution imagery, and severe geometric distortions near the panorama projection poles, the authors propose several solutions. Firstly, they leverage training examples from conventional datasets such as ImageNet and COCO to overcome the lack of annotated training data specific to equirectangular panoramas. Secondly, they employ low-resolution images that require only moderate GPU computing power and memory, making it feasible to process high-resolution equirectangular panoramas efficiently. The key innovation of their approach lies in the multi-projection YOLO detector which handles projection distortions by creating multiple stereographic sub-projections allowing for more accurate object detection even in regions with severe geometric distortions near the panorama projection poles. In their experiments, they compare their multi-projection YOLO detector with other state-of-the-art detectors like Faster RCNN and demonstrate that YOLO outperforms Faster RCNN in terms of accuracy for object detection in equirectangular panoramas. Furthermore, their multi-projection YOLO achieves the best accuracy when using low resolution input. Overall, this paper presents a comprehensive solution for object detection in equirectangular panoramas by addressing challenges related to training data availability, computational efficiency and handling geometric distortions. The proposed multi-projection YOLO detector shows promising results and opens up new possibilities for applications involving high resolution virtual reality imagery.