PaperlayerView exact revision on arXiv

Source-linked AI summary

UnitBox: An Advanced Object Detection Network

Jiahui Yu, Yuning Jiang, Zhangyang Wang, Zhimin Cao, Thomas Huang

arXiv:1608.01471v1cs.CV

TL;DR

Existing CNN detectors often treat bounding-box bounds as independent variables and face efficiency and scale-variation challenges. UnitBox combines fully convolutional prediction with an IoU loss that jointly regresses the box as a unit. It reports more accurate localization, faster convergence, variable-scale capability, and the best published FDDB face-detection performance.

  • Problem

    Existing methods independently regress correlated bounding-box bounds with ℓ2 loss, while proposal and image-pyramid procedures limit robustness or efficiency.

  • Method

    UnitBox uses a fully convolutional network and a novel IoU loss that directly optimizes overlap while jointly regressing all box bounds.

  • Results

    UnitBox achieves more accurate localization, faster convergence, variable-scale training and single-scale testing, and the best published FDDB detection performance.

  • Takeaways & Limitations

    The IoU loss and UnitBox provide a box-level localization approach applicable to object localization and detection tasks.

Abstract

from arXiv · show

In present object detection systems, the deep convolutional neural networks (CNNs) are utilized to predict bounding boxes of object candidates, and have gained performance advantages over the traditional region proposal methods. However, existing deep CNN methods assume the object bounds to be four independent variables, which could be regressed by the $\ell_2$ loss separately. Such an oversimplified assumption is contrary to the well-received observation, that those variables are correlated, resulting to less accurate localization. To address the issue, we firstly introduce a novel Intersection over Union ($IoU$) loss function for bounding box prediction, which regresses the four bounds of a predicted box as a whole unit. By taking the advantages of $IoU$ loss and deep fully convolutional networks, the UnitBox is introduced, which performs accurate and efficient localization, shows robust to objects of varied shapes and scales, and converges fast. We apply UnitBox on face detection task and achieve the best performance among all published methods on the FDDB benchmark.

1. INTRODUCTION

CNN-based detectors improve object detection but face efficiency, scale-variation, and bounding-box localization challenges. UnitBox addresses these issues with fully convolutional prediction and an IoU loss that jointly regresses box bounds.

  • Existing detection pipelines: CNN-based detection pipelines use region proposals, CNN recognition, and bounding-box regression, but proposal extraction can be fragile and computationally expensive.Proposal methods may fail under local appearance changes such as partial occlusion and can hinder real-time deployment.
  • Existing detection pipelines: Faster R-CNN relies on fixed anchor-box scales and aspect ratios, making large shape variations and especially small objects difficult to handle.
  • DenseBox limitations: DenseBox regresses four pixel-to-bound distances independently with ℓ2 loss, despite correlations among the object bounds.This can produce boxes where some bounds are accurate but the complete predicted box is unacceptable.
  • DenseBox limitations: DenseBox requires fixed-scale training patches and image pyramids, which reduce detection efficiency across varied object scales.
  • UnitBox: UnitBox uses a fully convolutional network to predict object bounds and pixel-wise classification scores directly on feature maps.
  • UnitBox: Its IoU loss jointly regresses all bounds as a unit, improving localization accuracy and convergence while supporting varied object shapes and scales.The paper reports the best published performance on FDDB for face detection.

2. IOU LOSS LAYER

The IoU loss layer jointly regresses all four bounding-box bounds as one unit, addressing the limitations of independent ℓ2 optimization. Its scale-normalized formulation supports multi-scale training and single-scale testing.

  • Limitations of ℓ2 Loss: ℓ2 loss optimizes the four bounding-box coordinates independently, despite their correlation, causing cases where individually close bounds still form unacceptable boxes.The paper identifies this as a major drawback of conventional bounding-box regression.
  • IoU Loss Forward: The IoU loss computes localization error from the predicted and ground-truth boxes' intersection and union areas.The forward algorithm defines predicted area, ground-truth area, intersection dimensions, intersection area, and union area.
  • IoU Loss Forward: L = −ln(IoU) treats the bounding box as a unit rather than optimizing its four coordinates independently.The loss directly favors larger intersection and a smaller predicted box, with a perfect match when the intersection equals the predicted box.
  • Scale Handling: IoU is naturally normalized to [0, 1], enabling multi-scale training and testing on a single-scale image.This normalization removes the need to balance bounding boxes solely through fixed-scale training patches.
  • IoU Loss Backward: The backward computation separates the predicted-box area penalty from the intersection-area penalty when deriving the localization-loss gradient.The gradient encourages maximizing intersection while minimizing predicted-box area.

3. UNITBOX NETWORK

UnitBox is a fully convolutional network that predicts pixel-wise object confidence and bounding boxes directly from feature maps. Its localization procedure uses confidence heatmaps to select candidate pixels and corresponding box predictions.

  • Architecture: UnitBox removes VGG-16 fully connected layers and adds fully convolutional branches for pixel-wise bounding-box and classification-score prediction.The network is derived from VGG-16 and predicts both outputs directly on feature maps.
  • Confidence Branch: The confidence branch produces a full-image one-channel heatmap trained with sigmoid cross-entropy.It uses convolution, up-sampling, and cropping to align the feature map with the input image.
  • Branch Design: The bounding-box branch is placed later than the confidence branch because unit regression requires a larger receptive field.The architecture uses stage 4 for confidence and stage 5 for bounding-box prediction.
  • Localization: For face detection, UnitBox fits ellipses to thresholded confidence heatmaps, then extracts boxes from selected center pixels.The ellipse stage supplies coarse face locations before the corresponding pixel-wise box predictions refine localization.

4. EXPERIMENTS

Experiments on FDDB show that IoU loss trains UnitBox more quickly and steadily than ℓ2 loss, handles scale variation better, and achieves the best published FDDB result. UnitBox also runs at about 12 fps on VGA images.

  • 4.1 Effectiveness of IoU Loss: UnitBox with IoU loss converges more quickly and steadily than UnitBox-ℓ2, maintaining a much lower miss rate during fine-tuning.The comparison keeps the network architecture and other parameters unchanged while reducing the ℓ2-loss learning rate to 10^-13.
  • 4.1 Effectiveness of IoU Loss: With fewer iterations, the best UnitBox model at approximately 16k iterations significantly outperforms UnitBox-ℓ2 at approximately 29k iterations on ROC curves.The models are selected at their respective best iterations for comparison.
  • 4.1 Effectiveness of IoU Loss: Across testing images resized from 60 to 960 pixels, IoU loss handles varied object scales better than ℓ2 loss.In the 960-pixel extreme case, UnitBox still gives a reasonable bounding box even when the face exceeds its approximately 200-pixel receptive field, whereas UnitBox-ℓ2 fails.
  • 4.2 Performance of UnitBox: UnitBox achieves the best detection result on FDDB among all published methods.The comparison uses ROC curves and an unshared UnitBox detector trained to further improve detection performance.
  • 4.2 Performance of UnitBox: UnitBox processes VGA-size images at about 12 fps, whereas DenseBox needs seconds per image.The authors identify this efficiency advantage as making UnitBox potential for real-time detection systems.

5. CONCLUSIONS

The paper concludes that IoU loss jointly regresses bounding-box bounds, improving localization accuracy and convergence over ℓ2 loss. UnitBox builds on this loss for state-of-the-art face detection on FDDB and may extend to other localization and detection tasks.

  • 5. CONCLUSIONS: IoU loss regresses an object candidate’s bounding box as a whole unit rather than as four independent variables.The loss is proposed for bounding-box prediction and is contrasted with the ℓ2 loss used in previous work.
  • 5. CONCLUSIONS: Compared with ℓ2 loss, IoU loss produces faster convergence and more accurate object localization.These are the paper’s stated effects of jointly optimizing the correlated bounds.
  • 5. CONCLUSIONS: UnitBox applies IoU loss in an advanced object detection network and achieves state-of-the-art performance on face detection.The reported application is the FDDB face-detection task.
  • 5. CONCLUSIONS: The authors propose that IoU loss and UnitBox could be valuable for other object localization and detection tasks.This is presented as the authors’ stated broader potential value, not as an evaluated result in the supplied passage.
Loading 1608.01471v1…