PaperlayerView exact revision on arXiv

Source-linked AI summary

Region Proposal by Guided Anchoring

Jiaqi Wang, Kai Chen, Shuo Yang, Chen Change Loy, Dahua Lin

arXiv:1901.03278v2cs.CV

TL;DR

Modern detectors rely on dense, uniform anchors with predefined shapes, creating efficiency and prior-design challenges. This paper introduces Guided Anchoring, which uses semantic features to predict likely object locations and location-dependent anchor shapes, with feature adaptation for those shapes. It reports higher recall with fewer anchors and consistent mAP gains across multiple anchor-based detectors.

  • Problem

    Uniform predefined anchors can require hand-picked scales and aspect ratios, while maintaining proposal recall demands many computationally costly anchors that are mostly irrelevant.

  • Method

    Guided Anchoring jointly predicts likely object-center locations and location-dependent anchor shapes from semantic features, then adapts features to the predicted geometry.

  • Results

    Guided Anchoring achieves 9.1% higher recall with 90% fewer anchors than the RPN baseline and improves detector mAP by up to 2.7%.

  • Takeaways & Limitations

    The scheme generates non-uniform arbitrary-shaped anchors and transfers across proposal methods and anchor-based detectors.

Abstract

from arXiv · show

Region anchors are the cornerstone of modern object detection techniques. State-of-the-art detectors mostly rely on a dense anchoring scheme, where anchors are sampled uniformly over the spatial domain with a predefined set of scales and aspect ratios. In this paper, we revisit this foundational stage. Our study shows that it can be done much more effectively and efficiently. Specifically, we present an alternative scheme, named Guided Anchoring, which leverages semantic features to guide the anchoring. The proposed method jointly predicts the locations where the center of objects of interest are likely to exist as well as the scales and aspect ratios at different locations. On top of predicted anchor shapes, we mitigate the feature inconsistency with a feature adaption module. We also study the use of high-quality proposals to improve detection performance. The anchoring scheme can be seamlessly integrated into proposal methods and detectors. With Guided Anchoring, we achieve 9.1% higher recall on MS COCO with 90% fewer anchors than the RPN baseline. We also adopt Guided Anchoring in Fast R-CNN, Faster R-CNN and RetinaNet, respectively improving the detection mAP by 2.2%, 2.7% and 1.2%. Code will be available at https://github.com/open-mmlab/mmdetection.

1. Introduction

Modern detectors commonly use dense, uniform anchors with predefined scales and aspect ratios, but this can require many irrelevant anchors and hand-picked priors. Guided Anchoring instead predicts likely object locations and location-dependent shapes, while adapting features to anchor geometry.

  • 1. Introduction: Dense anchoring places k predefined-scale and aspect-ratio anchors at every feature-map location.The scheme follows alignment and consistency rules, matching anchor centers to feature pixels and receptive fields to anchor scale and shape.
  • 1. Introduction: Uniform fixed-shape anchors require problem-specific priors and can reduce speed and accuracy when designed poorly.Maintaining proposal recall also requires many anchors, most of which are irrelevant object candidates and increase computational cost.
  • 1. Introduction: Guided Anchoring generates sparse anchors by first identifying object-containing sub-regions and then predicting shapes at selected locations.The method is motivated by non-uniform object locations and the relationship between object scale, image content, location, and scene geometry.
  • 1. Introduction: Learnable anchor shapes create feature-consistency challenges because scales and aspect ratios vary across feature-map locations.The proposed feature adaptation module adapts representations according to anchor geometry.
  • 1. Introduction: Guided Anchoring improves recall by 9.1% with 90% fewer anchors than the dense RPN baseline.Integrated into Fast R-CNN, Faster R-CNN, and RetinaNet, it improves overall mAP by 2.2%, 2.7%, and 1.2%, respectively, on COCO.
  • 1. Introduction: The method predicts non-uniform, arbitrary-shaped anchors and studies high-quality proposals for improving trained two-stage detectors.Its contributions include factorized anchor-distribution modeling, feature adaptation, and a fine-tuning scheme based on GA-RPN proposals.

2. Related Work

Prior anchor-based detectors generally use dense, uniform sliding-window anchors across feature maps. Guided Anchoring differs by discarding this scheme in favor of sparse, semantically guided anchors.

  • 2. Related Work: Faster R-CNN introduced an RPN that maps sliding-window anchors to features for generating object proposals.Later two-stage methods adopted the same general sliding-window anchoring design.
  • 2. Related Work: Single-stage detectors retain densely generated sliding-window anchors even when they omit a separate proposal stage.SSD and DenseBox evaluate dense anchors like a multiclass RPN, while RetinaNet uses focal loss for foreground-background imbalance.
  • 2. Related Work: Unlike prior single-, two-, and multi-stage methods, Guided Anchoring replaces dense uniform sliding-window anchors with sparse, guided anchors.The cited comparison identifies this as the method’s primary difference from conventional detectors.

3. Guided Anchoring

Guided Anchoring replaces dense, uniform, predefined anchors with sparse, image-guided anchors whose locations and shapes are predicted from features. It also adapts features to anchor geometry to preserve alignment and consistency.

  • Anchor generation: The anchor-generation module factorizes object distributions into location prediction and shape prediction conditioned on location.The location branch identifies likely object centers, while the shape branch predicts the most probable shape at each selected location.
  • Multi-level generation: Multi-level anchor generation collects anchors from feature maps at different pyramid levels while sharing parameters across levels.The scheme follows the FPN architecture and is parameter-efficient because its anchor-generation parameters are shared.
  • Anchor generation: 90% fewer anchors while maintaining recall is enabled by filtering locations with predicted object-center probabilities.The location branch produces a probability map and retains locations above a threshold; this can filter out 90% of regions while maintaining the same recall.
  • Anchor generation: Guided Anchoring predicts location-dependent anchor shapes instead of assigning predefined shapes at every feature-map location.The shape branch predicts width and height for each selected location, allowing arbitrary aspect ratios and better coverage of tall or wide objects.
  • Anchor-guided feature adaptation: Feature adaptation transforms each location’s feature according to its predicted anchor shape to address representation inconsistency.A deformable convolution uses offsets predicted from the shape branch to adapt the original feature map at each location.
  • Training: Training assigns center, ignore, and outside regions around ground-truth boxes and matches variable-shape anchors using an approximated variable-anchor IoU.The variable-anchor IoU is approximated by sampling nine width-height pairs and taking the maximum IoU.

4. Experiments

Experiments on MS COCO evaluate Guided Anchoring for region proposals, detection, fine-tuning, and component effectiveness. The results show higher recall and detection performance, while feature adaptation and sparse anchor selection support the method’s effectiveness and efficiency.

  • Region proposal results: Guided Anchoring improves AR300 by 10.5% and AR1000 by 9.1% over the RPN baseline on MS COCO.
  • Detection results: Guided Anchoring improves detection mAP by 2.3%, 2.7% and 1.2% in Fast R-CNN, Faster R-CNN and RetinaNet, respectively.
  • Fine-tuning results: High-quality GA-RPN proposals improve a trained Faster R-CNN by 2.2% mAP after three fine-tuning epochs, whereas RPN proposals provide no gain.
  • Ablation study: The location threshold can reduce the number of anchors by more than 90% with only a minor recall decrease, supporting more efficient inference.
  • Ablation study: Feature adaptation substantially improves recall by rearranging features according to predicted anchor shapes, while simply adding deformable convolution performs worse.
  • Ablation study: Predicted anchors cover a wider range of scales and aspect ratios than predefined sliding-window anchors, with distributions similar to ground-truth objects.

5. Conclusion

Guided Anchoring replaces uniform dense anchors with semantic, location-dependent anchors and improves proposal recall while reducing anchor count. It also transfers across anchor-based detectors, improving performance by as much as 2.7%.

  • Guided Anchoring leverages semantic features to generate non-uniform anchors with location-dependent shapes.
  • 9.1% higher recall is achieved with 90% fewer anchors than the RPN baseline using sliding-window anchors.
  • Up to 2.7% performance improvement is obtained when Guided Anchoring is applied to various anchor-based detectors.

RPN GA - RPN

The figure compares proposals generated by the RPN and GA-RPN methods. RPN proposals appear in the top row, while GA-RPN proposals appear in the bottom row.

  • The top row shows examples of RPN proposals.
  • The bottom row shows examples of GA-RPN proposals.
  • The figure is organized as a visual comparison between RPN and GA-RPN proposals.
Loading 1901.03278v2…