# Learning Objectness from Sonar Images for Class-Independent Object Detection

**Authors:** Matias Valdenegro-Toro  
**arXiv:** [1907.00734](https://arxiv.org/abs/1907.00734)  
**Date:** 2019-07-01  
**Summary page:** https://www.summarizepaper.com/en/paper/1907.00734v1/

## Key points

- The paper addresses the challenge of detecting novel objects without class information in sonar images
- Proposes a fully convolutional neural network that directly regresses an objectness value from a Forward-Looking sonar image
- Achieves high recall (96%) with only 100 proposals per image, outperforming other methods like EdgeBoxes and Selective Search
- Demonstrates the ability to generalize to completely new objects and potentially valuable in real-world applications such as finding lost objects under the sea
- Offers promising results and potential applications in underwater robotics and related fields

## Summary

The paper titled "Learning Objectness from Sonar Images for Class-Independent Object Detection" addresses the challenging task of detecting novel objects without class information, particularly in the context of underwater robotics. Modeling marine objects in sonar images is inherently difficult, and training data may not be readily available. While detection proposal algorithms can be used for this purpose, they typically require a large number of output bounding boxes. To overcome these challenges, the authors propose a fully convolutional neural network that directly regresses an objectness value from a Forward-Looking sonar image. By ranking objectness, the network achieves high recall (96%) with only 100 proposals per image. In comparison, other methods like EdgeBoxes require 5000 proposals to achieve slightly better recall (97%), while Selective Search requires 2000 proposals to achieve 95% recall. Furthermore, the authors demonstrate that their method outperforms a template matching baseline by a significant margin and has the ability to generalize to completely new objects. This suggests that their technique could be valuable in real-world applications such as finding lost objects under the sea. Overall, this paper presents an innovative approach to address the challenge of detecting novel objects without class information in sonar images. The proposed fully convolutional neural network offers promising results and potential applications in underwater robotics and related fields.