PaperlayerView exact revision on arXiv

Source-linked AI summary

Grid R-CNN

Xin Lu, Buyu Li, Yuxin Yue, Quanquan Li, Junjie Yan

arXiv:1811.12030v1cs.CV

TL;DR

Object detectors commonly localize boxes through offset regression, motivating a method that preserves explicit spatial information and improves difficult localization. Grid R-CNN predicts multiple grid points with a fully convolutional branch and fuses neighboring point features; on COCO, it reports 4.1% and 10.0% AP gains at IoU=0.8 and 0.9 against Faster R-CNN with ResNet-50 and FPN.

  • Problem

    Traditional bounding-box localization typically uses regression branches that collapse high-level feature maps into vectors, limiting explicit spatial representation.

  • Method

    Grid R-CNN replaces offset regression with fully convolutional grid-point prediction, multi-point supervision, and feature-map-level fusion of neighboring grid points.

  • Results

    4.1% AP gain at IoU=0.8 and 10.0% AP gain at IoU=0.9 are reported on COCO compared to Faster R-CNN with ResNet-50 and FPN.

  • Takeaways & Limitations

    Grid R-CNN achieves high-quality object localization and provides consistent improvements across detection frameworks and network architectures.

  • Takeaways & Limitations

    Simply enlarging the proposal area introduces redundant background or other-object features and harms small-object detection accuracy.

Abstract

from arXiv · show

This paper proposes a novel object detection framework named Grid R-CNN, which adopts a grid guided localization mechanism for accurate object detection. Different from the traditional regression based methods, the Grid R-CNN captures the spatial information explicitly and enjoys the position sensitive property of fully convolutional architecture. Instead of using only two independent points, we design a multi-point supervision formulation to encode more clues in order to reduce the impact of inaccurate prediction of specific points. To take the full advantage of the correlation of points in a grid, we propose a two-stage information fusion strategy to fuse feature maps of neighbor grid points. The grid guided localization approach is easy to be extended to different state-of-the-art detection frameworks. Grid R-CNN leads to high quality object localization, and experiments demonstrate that it achieves a 4.1% AP gain at IoU=0.8 and a 10.0% AP gain at IoU=0.9 on COCO benchmark compared to Faster R-CNN with Res50 backbone and FPN architecture.

1. Introduction

Grid R-CNN replaces offset regression with grid-guided, fully convolutional localization that preserves spatial information. Multi-point supervision and feature-map fusion improve robustness and localization quality, with gains especially at strict IoU thresholds.

  • 1. Introduction: 4.1% AP gain at IoU=0.8 and 10.0% AP gain at IoU=0.9 are reported against Faster R-CNN with ResNet-50 and FPN on COCO.These results indicate stronger high-quality object localization under stricter overlap criteria.
  • 1. Introduction: Grid R-CNN replaces traditional regression with a fully convolutional grid-guided localization framework that preserves spatial information.The method predicts grid-point locations on the pixel level to determine the object bounding box.
  • 1. Introduction: Multi-point supervision reduces the impact of inaccurate individual point predictions by providing additional spatial clues.A grid allows neighboring or boundary points to calibrate potentially inaccurate predictions.
  • 1. Introduction: Feature-map-level fusion incorporates complementary information from spatially related grid points to calibrate their locations.Neighboring grid-point feature maps are fused so each point benefits from correlated spatial features.
  • 1. Introduction: Grid R-CNN is designed as a broadly applicable alternative to regression-based localization across detection frameworks and architectures.The authors report consistent gains across different frameworks and network architectures.

2. Related Works

Related two-stage detectors extract RoI features for classification and localization, while other methods use fully convolutional or keypoint-based designs. Grid R-CNN follows a top-down two-stage pipeline but replaces box offset regression with grid-guided localization.

  • 2. Related Works: Two-stage detectors generate region proposals and extract shared feature representations for classifying and localizing each RoI.The related-work discussion traces this design from R-CNN through later Faster R-CNN extensions.
  • 2. Related Works: R-FCN replaces the original fully connected network with a region-based fully convolutional network, while FPN builds multi-scale semantic feature maps.Mask R-CNN adds a parallel pixel-wise object-mask branch to Faster R-CNN.
  • 2. Related Works: CornerNet is a one-stage bottom-up detector that detects and groups paired keypoints to form object bounding boxes.Its keypoints are generated from the entire image without initially defining an instance.
  • 2. Related Works: Grid R-CNN instead uses a top-down two-stage approach that defines instances first and then localizes their bounding boxes with grid points.Its localization branch operates on RoI features and uses a fully convolutional grid prediction mechanism.

3. Grid R-CNN

Grid R-CNN replaces offset regression with fully convolutional grid-point localization, then improves point prediction through spatial feature fusion and extended heatmap mapping. The resulting design addresses incomplete proposal coverage while preserving spatial representations.

  • Grid Guided Localization: Grid R-CNN replaces fully connected offset regression with a fully convolutional branch that predicts predefined grid-point locations.The branch preserves spatial information and produces heatmaps for grid-point localization.
  • Grid Guided Localization: A fixed N × N grid aligns target points with the object box, and each point is supervised through a heatmap with positive pixel locations.RoIAlign features are processed by dilated convolutions and deconvolution layers before producing N × N heatmaps.
  • Grid Guided Localization: The four box boundaries are computed from probability-weighted coordinates of grid points assigned to the corresponding edges.For example, the upper boundary is the probability-weighted average of the three upper grid points.
  • Extended Region Mapping: Extended region mapping doubles the heatmap’s represented image region without enlarging the RoI feature-extraction area, covering target grid points for positive proposals in most cases.This addresses proposals whose heatmaps fail to cover ground-truth grid points; after remapping, all target points of positive proposals are covered.
  • Extended Region Mapping: Simply enlarging proposal areas introduces redundant background or object features and brings no gain while harming small-object detection accuracy.The extended mapping strategy instead changes the heatmap-to-image relationship while keeping RoI extraction unchanged.

4. Experiments

Experiments across Pascal VOC and COCO evaluate Grid R-CNN’s components, generalization, and localization quality. Ablations show gains from multi-point supervision, feature fusion, and extended region mapping, while comparisons report stronger performance than regression-based baselines, especially at high IoU thresholds.

  • Experimental Setup: Experiments span Pascal VOC and COCO, including minival and test-dev evaluations across detection frameworks and backbone architectures.The COCO experiments use 80 object categories and compare multiple framework and backbone configurations.
  • Ablation Study: Increasing supervised grid points improves detection accuracy, supporting the multi-point supervision design.The ablation compares 2-point, 4-point, and 9-point grid formulations without feature fusion.
  • Ablation Study: 0.7% AP is gained by complete second-order feature fusion over the non-fusion baseline, with larger improvement at AP0.75 than AP0.5.The result indicates that feature fusion improves bounding-box localization accuracy.
  • Ablation Study: 1.2% AP is gained by extended region mapping, which improves APlarge without negatively affecting APsmall.Direct proposal enlargement improves APlarge but decreases APsmall, whereas extended mapping avoids that trade-off.
  • Comparison with State-of-the-art Methods: 3.6% and 9.7% AP improvements over FPN and R-FCN, respectively, are reported on Pascal VOC under a COCO-style criterion.The criterion averages AP across IoU thresholds from 0.5 to 0.95.
  • Analysis and Discussion: 4.1% and 10% gains over the baseline at AP0.8 and AP0.9 show that Grid R-CNN’s main advantage is higher-quality bounding-box localization.Grid R-CNN outperforms regression at IoU thresholds greater than 0.7.

5. Conclusion

Grid R-CNN replaces box-offset regression with grid-guided localization, using FCN position sensitivity, feature fusion, and extended region mapping for high-quality detection. Experiments show consistent improvements, especially under strict IoU metrics, while the approach remains extensible to other frameworks.

  • Grid R-CNN replaces traditional box-offset regression with a grid-guided mechanism for high-quality localization.The framework predicts grid points using FCN position sensitivity and determines the bounding box from the grid.
  • Feature fusion calibrates grid-point locations by transferring spatial information at the feature-map level.
  • Extended region mapping enlarges RoI representational coverage so more grid points are included, significantly improving performance.
  • Grid R-CNN delivers solid, consistent improvements and state-of-the-art performance, particularly at AP evaluated with IoU=0.8 and IoU=0.9.
  • The grid-guided localization approach is easy to extend to other detection frameworks, including possible combinations with scale selection and cascade techniques.
Loading 1811.12030v1…