PaperlayerView exact revision on arXiv

Source-linked AI summary

Corner Proposal Network for Anchor-free, Two-stage Object Detection

Kaiwen Duan, Lingxi Xie, Honggang Qi, Song Bai, Qingming Huang, Qi Tian

arXiv:2007.13816v1cs.CV

TL;DR

Object detection must locate objects and identify their classes despite wide variation in their geometry and appearance. CPN extracts corner-combination proposals and classifies them in two steps; it reports strong MS-COCO accuracy and efficient operating points.

  • Problem

    Objects vary in scale, shape, and position, making detection a challenge of locating varied geometries and assigning accurate labels.

    Source: §1 · pp. 1–2

  • Method

    CPN forms object proposals by enumerating valid top-left and bottom-right corner combinations, then applies binary filtering and multi-class classification.

    Source: §1 · p. 2

  • Results

    49.2% AP on MS-COCO with a 104-layer Hourglass backbone outperforms CenterNet by 2.2%.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Anchor-free proposals offer flexibility for locating varied object geometries, while an individual discrimination stage improves precision and a properly implemented two-stage framework can be efficient.

    Source: §5 · p. 15

Abstract

from arXiv · show

The goal of object detection is to determine the class and location of objects in an image. This paper proposes a novel anchor-free, two-stage framework which first extracts a number of object proposals by finding potential corner keypoint combinations and then assigns a class label to each proposal by a standalone classification stage. We demonstrate that these two stages are effective solutions for improving recall and precision, respectively, and they can be integrated into an end-to-end network. Our approach, dubbed Corner Proposal Network (CPN), enjoys the ability to detect objects of various scales and also avoids being confused by a large number of false-positive proposals. On the MS-COCO dataset, CPN achieves an AP of 49.2% which is competitive among state-of-the-art object detection methods. CPN also fits the scenario of computational efficiency, which achieves an AP of 41.6%/39.7% at 26.2/43.3 FPS, surpassing most competitors with the same inference speed. Code is available at https://github.com/Duankaiwen/CPNDet

1 Introduction

Object detection must handle objects with varied scales, shapes, and positions while locating them and assigning accurate labels. CPN combines anchor-free corner proposals with standalone classification, and reports strong accuracy and speed on MS-COCO.

1 Introduction

  • Detection is challenging because objects vary in scale, shape, and position, and members of the same class can look very different.The paper frames finding varied geometries and assigning accurate labels as the two keys to detection.Source: §1 · p. 1
  • The paper argues anchor-free border-based methods can better recall rare geometries, while an individual classifier addresses their false-positive proposals.The authors frame the design as combining anchor-free proposal extraction with two-stage discrimination.Source: §1 · p. 2
  • CPN locates top-left and bottom-right corners, enumerates valid combinations as proposals, then uses binary filtering and stronger multi-class classification to reject false pairs and label survivors.The proposal strategy replaces CornerNet's keypoint feature embedding for pairing corners.Source: §1 · p. 2
  • 49.2% AP on MS-COCO with a 104-layer Hourglass backbone outperforms CenterNet by 2.2%, while lighter settings reach 41.6% at 26.2 FPS or 39.7% at 43.3 FPS.The lighter settings use DLA-34 and switch off image flip during inference.Source: §1 · p. 3

2 Related Work

Related work is organized around how detectors locate objects and determine their classes. The section contrasts fixed-anchor methods, keypoint-based anchor-free methods, and alternative classification strategies.

2 Related Work

  • Anchor-based detectors place fixed-scale, fixed-shape regional proposals uniformly across the image, then classify proposals for objectness and class.Related work improves regional features or aligns proposals and features.Source: §2 · p. 3
  • Anchor-free methods use one or a few keypoints; point-grouping detectors combine keypoints, while point-vector detectors pair one keypoint with object geometry.The cited examples include CornerNet and CenterNet for grouping, and FCOS and FoveaBox for point-vector approaches.Source: §2 · p. 3
  • Two-stage detectors train an individual classifier for proposals, whereas one-stage detectors mostly use classification cues from an earlier stage; two-stage methods are often more accurate but slower.Prior work partitions classification into an initial easy-false-positive filter and a heavier classifier for surviving proposals.Source: §2 · p. 4

3 Our Approach

CPN pairs anchor-free corner-based proposal extraction with two-stage classification, aiming to retain flexibility for varied object geometry while filtering false positives. Its approach uses a lightweight objectness classifier before a class-label classifier.

3.1 Anchor-based or Anchor-free? One-stage or Two-stage?

  • Anchor-free point-grouping methods, including CPN, are presented as more flexible for unusual object geometries than fixed-anchor approaches, which can have lower recall for very large or extreme-aspect-ratio objects.The section highlights low recall for Faster R-CNN at aspect ratios of 5:1 and 8:1, and explains that distance-to-border prediction can also be difficult for far borders.Source: §3.1 · pp. 4–5

3.2 The Framework of Corner Proposal Network

  • CPN extracts proposals by pairing same-class top-left and bottom-right corner keypoints, leaving incorrectly paired proposals for the classification stage to reject.The method follows CornerNet for corner detection but enumerates valid combinations rather than grouping keypoints with learned embeddings.Source: §3.2 · pp. 6–7
  • A lightweight binary classifier first filters proposals, followed by a regional-feature classifier that assigns each surviving proposal a class label.The first classifier uses proposal features; the second uses features from the category feature map because corner-class consensus may be unreliable without ROI information.Source: §3.2 · p. 8
  • Compared with DeNet, CPN uses multi-class corner labels, a lightweight binary filter before class classification, and a focal-loss variant for proposal imbalance.The authors state that the lightweight first classifier helps efficiency and that the loss addresses the imbalance between positive and negative proposals.Source: §3.2 · pp. 8–9

3.3 The Inference Process

  • CPN uses a two-step inference process with a 0.2 first-stage threshold, about 20% of proposals surviving, and the two classifiers’ overheads described as comparable.The first classifier costs about 10% as much as the second per proposal; the method combines corner and regional classification scores and keeps the top 100 proposals.Source: §3.3 · p. 9

4 Experiments

Experiments evaluate CPN on MS-COCO using standard detection metrics and compare it with state-of-the-art detectors. The results report strong accuracy, fewer false positives, and competitive speed.

4.1 Dataset and Evaluation Metrics

  • Experiments use COCO trainval35k, report on test-dev, and use the remaining 5K validation images for ablations and visualizations; AP averages ten IoU thresholds.The dataset contains 120K images, over 1.5 million boxes, and 80 categories; evaluation preserves at most 100 proposals per image.Source: §4.1 · pp. 9–10

4.2 Implementation Details

  • Models use Hourglass-52/104 or DLA-34 backbones, 511×511 inputs, and 128×128 feature maps, with single- or multi-scale inference and soft-NMS.Training uses Adam on eight Tesla-V100 GPUs; the stated schedule is 200K iterations followed by 50K at a reduced learning rate.Source: §4.2 · p. 10

4.3 Comparisons with State-of-the-Art Detectors

  • 49.2% AP is CPN’s best reported result, surpassing published anchor-free approaches; CPN-104 also exceeds CenterNet by 2.1% single-scale and 2.2% multi-scale.CPN-52 reports 43.9% single-scale and 45.8% multi-scale AP, exceeding CornerNet-104 by 3.4% and 3.7%, respectively.Source: Table 3 · p. 11, §4.3 · pp. 11–12

4.4 Classification Improves Precision

  • CPN-52 and CPN-104 have average false discovery rates of 33.4% and 30.6%, respectively, lower than CornerNet and CenterNet.The table caption states that fewer false positives under the same corner keypoint extractor are key to outperforming the baselines in AP.Source: Table 5 · p. 13, §4.4 · p. 13
  • Combining the binary and multi-class classifiers improves AP by more than 1% beyond their separate absolute gains of 3.4% and 3.8%.The authors describe the classifiers as complementary because proposal objectness and class labeling require somewhat different semantic information.Source: §4.4 · pp. 12–13

4.5 Inference Speed

  • On MS-COCO validation using an NVIDIA Tesla-V100 GPU, CPN-104 achieves 7.3 FPS/46.8% AP versus CenterNet-104’s 5.1/44.8%; CPN-52 reaches 9.9 FPS/43.8% AP, outperforming CornerNet-52 and CenterNet-52.Source: §4.5 · p. 14

5 Conclusions

The conclusion presents an anchor-free, two-stage detector that combines corner-based proposal extraction with classification to filter false positives. It argues that flexible proposals and a separate discrimination stage improve recall and precision, and that this framework can be efficient.

5 Conclusions

  • The framework extracts corner-keypoint proposals, then uses two-step classification to filter false positives.Source: §5 · p. 15
  • The two stages significantly improve detection recall and precision, and the final result ranks among top existing methods.Source: §5 · p. 15
  • The authors conclude that anchor-free proposals are more flexible, while individual discrimination improves precision; a properly implemented two-stage framework can be efficient.They therefore regard the one-stage versus two-stage debate as not critical.Source: §5 · p. 15

Related papers

Loading 2007.13816v1…