One-Shot General Object Localization

Yang You, Zhuochen Miao, Kai Xiong, Weiming Wang, Cewu Lu

Abstract

This paper presents a general one-shot object localization algorithm called OneLoc. Current one-shot object localization or detection methods either rely on a slow exhaustive feature matching process or lack the ability to generalize to novel objects. In contrast, our proposed OneLoc algorithm efficiently finds the object center and bounding box size by a special voting scheme. To keep our method scale-invariant, only unit center offset directions and relative sizes are estimated. A novel dense equalized voting module is proposed to better locate small texture-less objects. Experiments show that the proposed method achieves state-of-the-art overall performance on two datasets: OnePose dataset and LINEMOD dataset. In addition, our method can also achieve one-shot multi-instance detection and non-rigid object localization. Code repository: https://github.com/qq456cvb/OneLoc.

Introduction

Standard object detection models, such as YOLO or Faster R-CNN, are highly effective but require extensive annotated datasets for every specific category. Recognizing a novel object typically necessitates time-consuming data collection and model retraining. "One-shot" localization addresses this by attempting to locate a new object in a complex scene using only a few reference images.

Existing approaches often suffer from two main drawbacks:

  1. Exhaustive Matching: These methods are computationally expensive and scale poorly as they compare query images against numerous reference templates.
  2. Generalization Issues: Deep learning models often overfit to training categories, failing on truly novel objects or those with minimal texture.

Methodology

OneLoc shifts the paradigm from simple template matching to a geometry-based voting approach. The process begins with feature extraction using SuperPoint descriptors from both reference and query images.

The OneLoc pipeline: from input image to predicted bounding box via center direction and size voting.

To maintain scale invariance, the network avoids predicting absolute pixel offsets, which vary with distance. Instead, it predicts unit center offset directions and relative sizes. During inference, the model employs a pairwise voting scheme where sampled points "vote" for the object's center. By calculating the intersection of these direction rays, the system analytically determines the most likely center. This approach is mathematically more robust than predicting bounding box corners, which typically exhibit higher variance.

Dense Equalized Voting

A significant challenge in localization is handling texture-less or smooth objects where traditional sparse "keypoints" (like corners or edges) are scarce. OneLoc addresses this through Dense Equalized Voting.

Comparison between sparse voting and the proposed dense equalized voting, showing reduced false positives.

Instead of relying on detected keypoints, the system samples points densely across the entire image using a stratified grid. This ensures that even feature-poor objects receive sufficient votes for accurate detection, while the "equalized" nature of the sampling prevents the model from being distracted by high-texture backgrounds.

Experimental Results

OneLoc was evaluated on the OnePose and LINEMOD benchmarks, as well as custom sets for multi-instance and non-rigid localization.

  • Accuracy: On the OnePose dataset, OneLoc achieved 87.80 mRec25 (mean recall at 25% IoU), outperforming the best baseline by nearly 20 points.
  • Robustness: On the challenging LINEMOD dataset, characterized by small, texture-less objects, it achieved 89.80 mRec25, significantly higher than the ~40.28 achieved by traditional feature-matching methods.
  • Efficiency: The system is highly performant, processing query images in approximately 133ms on a 1080Ti GPU, compared to over 2 seconds for matching-based baselines.
  • Versatility: OneLoc successfully localized non-rigid objects, such as crumpled clothing, where previous state-of-the-art methods like DTOID failed.

Conclusion

OneLoc represents a significant advancement in one-shot localization by combining scale-invariant geometric voting with dense sampling. While it currently focuses on 2D bounding boxes, the underlying logic provides a foundation for future extension into full 6D pose estimation. Future work will also look to improve performance under extreme occlusions and drastic viewpoint changes.

Code repository: https://github.com/qq456cvb/OneLoc.

Summary updated

Abstract

This paper presents a general one-shot object localization algorithm called OneLoc. Current one-shot object localization or detection methods either rely on a slow exhaustive feature matching process or lack the ability to generalize to novel objects. In contrast, our proposed OneLoc algorithm efficiently finds the object center and bounding box size by a special voting scheme. To keep our method scale-invariant, only unit center offset directions and relative sizes are estimated. A novel dense equalized voting module is proposed to better locate small texture-less objects. Experiments show that the proposed method achieves state-of-the-art overall performance on two datasets: OnePose dataset and LINEMOD dataset. In addition, our method can also achieve one-shot multi-instance detection and non-rigid object localization. Code repository: https://github.com/qq456cvb/OneLoc.
BibTeX
@misc{you2022one,
  author = {Yang You and Zhuochen Miao and Kai Xiong and Weiming Wang and Cewu Lu},
  title = {{One-Shot General Object Localization}},
  year = {2022},
  eprint = {2211.13392},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2211.13392},
}

Loading related content...

Figure