Acquisition of Localization Confidence
for Accurate Object Detection
Presentation by Haiyan Wang
Media Lab
08/24/2018
1
Outline
• Introduction
• Two problems
• IOU-Net
• PrRoI Pooling
• IoU-guided NMS
• Bounding box refinement as an optimization procedure
• Experiment result
2
Introduction
• Object detection
• Distinguish bg/fg
• NMS
(Non-maximum-suppression)
• Bounding box regression
(classification confidence
localization confidence = IOU)
3
Problem 1
• Accurate localized bounding boxes will be incorrectly eliminated in
the traditional NMS procedure.
4
• The correlation between the IOU of bounding boxes with the
matched ground-truth and the classification/localization confidence
• The correlation coefficients are: (a) 0.217, (b) 0.617
5
Problem 2
• Bounding box regression may degenerate the localization of input
bounding boxes If applied for multiple times.
6
• Non-monotonic bounding box regression
7
IOU-Net
8
• Jittered RoIs: manually transform them with a set of randomized
parameters, then remove from this candidate set the bounding boxes
having an IoU less than Ωtrain= 0.5 with the matched ground-truth
• The extracted features are fed into a two-layer network for the IOU
prediction.
9
PrRoI Pooling
10
• PrPool(Bin,F) is differentiable w.r.t. the coordinates of bin
• Since we avoids any quantization, PrPool is continuouslydifferentiable
11
IoU-guided NMS
12
• Use the predicted IoU instead of the classification confidence as the
ranking keyword for bounding boxes.
• Confidence Clustering, for a group of bounding boxes matchingthe
same ground-truth, we take the most confident prediction for the
class label
13
Precision
14
Recall
15
Bounding box refinement as an optimization
procedure
• Traditional bounding box refinement
• IoU-Net directly estimates , While the proposed
PreciseRoI Pooling layer enables the computation of the gradient of
IoU w.r.t. bounding box coordinates, we can directly use gradient
ascent method to find the optimal solution to Equation
16
17
Experiment result
18
19

Acquisition of Localization Confidence for Accurate Object Detection

  • 1.
    Acquisition of LocalizationConfidence for Accurate Object Detection Presentation by Haiyan Wang Media Lab 08/24/2018 1
  • 2.
    Outline • Introduction • Twoproblems • IOU-Net • PrRoI Pooling • IoU-guided NMS • Bounding box refinement as an optimization procedure • Experiment result 2
  • 3.
    Introduction • Object detection •Distinguish bg/fg • NMS (Non-maximum-suppression) • Bounding box regression (classification confidence localization confidence = IOU) 3
  • 4.
    Problem 1 • Accuratelocalized bounding boxes will be incorrectly eliminated in the traditional NMS procedure. 4
  • 5.
    • The correlationbetween the IOU of bounding boxes with the matched ground-truth and the classification/localization confidence • The correlation coefficients are: (a) 0.217, (b) 0.617 5
  • 6.
    Problem 2 • Boundingbox regression may degenerate the localization of input bounding boxes If applied for multiple times. 6
  • 7.
  • 8.
  • 9.
    • Jittered RoIs:manually transform them with a set of randomized parameters, then remove from this candidate set the bounding boxes having an IoU less than Ωtrain= 0.5 with the matched ground-truth • The extracted features are fed into a two-layer network for the IOU prediction. 9
  • 10.
  • 11.
    • PrPool(Bin,F) isdifferentiable w.r.t. the coordinates of bin • Since we avoids any quantization, PrPool is continuouslydifferentiable 11
  • 12.
    IoU-guided NMS 12 • Usethe predicted IoU instead of the classification confidence as the ranking keyword for bounding boxes. • Confidence Clustering, for a group of bounding boxes matchingthe same ground-truth, we take the most confident prediction for the class label
  • 13.
  • 14.
  • 15.
  • 16.
    Bounding box refinementas an optimization procedure • Traditional bounding box refinement • IoU-Net directly estimates , While the proposed PreciseRoI Pooling layer enables the computation of the gradient of IoU w.r.t. bounding box coordinates, we can directly use gradient ascent method to find the optimal solution to Equation 16
  • 17.
  • 18.
  • 19.

Editor's Notes

  • #14 Resolve the misalignment between the classification confidence and localization accuracy.
  • #17 Smooth L1 loss