# Recurrent Neural Networks for video object detection

**Authors:** Ahmad B Qasim, Arnd Pettirsch  
**arXiv:** [2010.15740](https://arxiv.org/abs/2010.15740)  
**Date:** 2020-10-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2010.15740v1/

## Key points

- Comparison of different methods for object detection in videos, specifically using Recurrent Neural Networks (RNNs)
- Inclusion of temporal context as a benefit in video object detection
- Conclusions and guidelines for video object detection networks
- Comparison of feature-based methods, box-level methods, and flow network methods
- Common outcomes among the compared methods, emphasizing the importance of incorporating temporal context
- Positive results from including RNNs in video object detection networks
- Results on YouTube Dataset and OTB Challenge Dataset showcasing performance of various architectures and models
- Proposed architecture includes region proposal network based on N-Gram concepts for detecting object bounding boxes within frames
- Attention mechanisms used to find saliency maps from deep feature maps obtained from SqueezeNet
- Attention module plays a role in obtaining input tensors for subsequent processing

## Summary

The existing summary discusses the comparison of different methods, specifically those using Recurrent Neural Networks (RNNs), for object detection in videos. It highlights the inclusion of temporal context as a benefit in video object detection and provides conclusions and guidelines for video object detection networks. Expanding on this, further details are provided. The study compares feature-based methods, box-level methods, and flow network methods for video object detection. Feature-based methods involve feeding feature maps from different frames into the recurrent units, while box-level methods feed bounding boxes with class probabilities into the recurrent units. Flow network methods utilize flow networks. The study indicates common outcomes among the compared methods, emphasizing the importance of incorporating temporal context in object detection. It also mentions that including RNNs in video object detection networks can yield positive results. Additionally, some specific findings are presented. Results on the YouTube Dataset and OTB Challenge Dataset are discussed, showcasing the performance of various architectures and models. The proposed architecture includes a region proposal network based on N-Gram concepts from Natural Language Processing to detect object bounding boxes within frames. Furthermore, attention mechanisms are used to find saliency maps from deep feature maps obtained from SqueezeNet. This attention module plays a role in obtaining input tensors for subsequent processing. Overall, this expanded summary provides a more detailed overview of the study's focus on comparing RNN-based methods for video object detection and its findings regarding different architectures and models used in the evaluation process such as YouTube Dataset and OTB Challenge Dataset which demonstrate improved performance when utilizing RNNs with attention modules for input tensor generation.