[Paper Review] Region Proposal by Guided Anchoring

Jiaqi Wang, Kai Chen+3 more|arXiv (Cornell University)|Jan 10, 2019

Advanced Neural Network ApplicationsComputer Science36 references59 citations

TL;DR

The paper introduces Guided Anchoring (GA) to predict location and shape of anchors jointly, plus an anchor-guided feature adaptation, achieving higher recall with far fewer anchors and improving detection across Fast R-CNN, Faster R-CNN, and RetinaNet.

ABSTRACT

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.

Motivation & Objective

• Motivate reducing reliance on dense, uniform anchors by predicting where objects may exist and what their shapes are at those locations.
• Develop a two-branch anchor prediction module (location and shape) to generate sparse, adaptive anchors.
• Introduce an anchor-guided feature adaptation to align features with location-specific anchor shapes.
• Investigate how high-quality GA proposals can improve two-stage and single-stage detectors and provide a practical fine-tuning workflow.

Proposed method

• Propose a two-branch anchor generation module that predicts object center locations (p(x,y|FI)) and location-dependent shapes (w,h) for anchors.
• Predict shapes as w=σ·s·exp(dw) and h=σ·s·exp(dh) to stabilize learning with a transform.
• Use a multi-level (FPN) implementation with shared anchor-generation parameters across levels.
• Apply an anchor-guided feature adaptation via a deformable convolution to align features with the predicted anchor shapes.
• Train with a multi-task loss including anchor localization loss (L_loc) and anchor shape loss (L_shape) in addition to standard classification and regression losses (L_cls, L_reg).
• Showcase a use-case where GA-RPN provides higher recall with 90% fewer anchors than RPN, and demonstrate improvements when GA is embedded in Fast R-CNN, Faster R-CNN, and RetinaNet.

Experimental results

Research questions

• RQ1 Can anchors be effectively generated in a sparse, non-uniform manner by jointly predicting object centers and shapes?
• RQ2 Does aligning features to the predicted anchor shapes via a dedicated adaptation module improve detection accuracy?
• RQ3 How do GA-based proposals influence two-stage and single-stage detectors compared to traditional dense anchors?
• RQ4 Can high-quality GA proposals enhance training and fine-tuning of existing detectors?

Key findings

• GA-RPN achieves 9.1% higher recall with 90% fewer anchors than the dense RPN baseline on COCO.
• Replacing anchors with GA improves mAP by 2.2% (GA-Fast-RCNN), 2.7% (GA-Faster-RCNN), and 1.2% (GA-RetinaNet).
• Anchor location and shape predictions, along with feature adaptation, yield significant recall gains and efficiency benefits.
• A fine-tuning scheme using GA-RPN proposals can boost a trained Faster R-CNN from 37.4% to 39.6% AP in 3 epochs.
• Using high-quality GA proposals with higher IoU thresholds reduces the need for many proposals and improves detector performance when training with fewer samples.