# Learning Open-World Object Proposals without Learning to Classify

**Authors:** Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In So Kweon, Weicheng Kuo  
**arXiv:** [2108.06753](https://arxiv.org/abs/2108.06753)  
**Date:** 2021-08-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.06753v1/

## Key points

- Authors address limitations of existing methods in handling novel objects in the field
- Binary classifiers tend to overfit to training categories, leading to poor generalization
- Proposed method estimates objectness of each region based on overlap with ground-truth objects in terms of location and shape
- OLN approach learns a simple strategy for estimating objectness without relying on classification
- OLN outperforms existing proposals on cross-category generalization using COCO dataset and cross-dataset evaluation on RoboNet, Object365, and EpicKitchens datasets
- OLN shows merit for long-tail object detection on LVIS dataset, improving performance for both rare and common categories
- Visualizations comparing proposed method with Mask R-CNN on COCO → EpicKitchens generalization are provided
- Results of long-tail detection on LVIS v0.5 using Faster R-CNN show improved performance with OLN-RPN replacement
- Overall, the paper introduces a novel approach to object proposal generation focusing on generalizable objectness estimation without relying on classification

## Summary

, , , , 


In this paper, the authors address the limitations of existing  in handling novel objects in the . They identify that binary classifiers in these methods tend to overfit to the training categories, leading to poor generalization. To overcome this issue, they propose a  that estimates the objectness of each region based on its overlap with ground-truth objects in terms of location and shape. The OLN approach learns a simple strategy for estimating objectness without relying on classification. The authors demonstrate that OLN outperforms existing proposals on cross-category generalization using COCO dataset and cross-dataset evaluation on RoboNet, Object365, and EpicKitchens datasets. They also show the merit of OLN for long-tail object detection on the large vocabulary dataset LVIS, where they observe clear improvements in both rare and common categories. Furthermore, the authors provide visualizations comparing their proposed method with Mask R-CNN on COCO → EpicKitchens generalization. They also present results of long-tail detection on LVIS v0.5 using Faster R-CNN, showing improved performance with OLN-RPN replacement. Overall, this paper introduces a novel approach to object proposal generation by focusing on generalizable objectness estimation without relying on classification. The proposed OLN method demonstrates superior performance compared to existing proposals across different datasets and shows promise for improving long-tail object detection tasks.