CenterNet

本資料は2020年8月28日に社内共有資料として展開していたものをWEBページ向けにリニューアルした内容になります。

■Purpose

Purpose of this material

• Understand an anchor free approach object detection algorithm

■Agenda

• Current object detection approaches
• Centernet approach
• Object as Points
• Training
  • Keypoint heatmap
  • Local offset
  • Size prediction
  • Loss function
• Network Architecture
  • DLA
  • Modified DLA
• Inference
• Results

■Background

Current approaches

• Object detections model (such as Yolo, SSD, etc.) rely on the usage of anchor boxes
• Anchor boxes are not completely optimal:
  • Wasteful: SSD300 does 8732 detections per class, and yolo448 does 98 detections per class, which means that most of the box are discarded
  • Inefficient: We have to process all the boxes (even we will discard them later), which comes with more processing time
  • Require post processing: like non-max suppression algorithm
  • Fixed: SSD requires fixed scale and steps of boxes, while yolov3 fixes the size of the anchors per detection level

■Centernet

Centernet approach

• End-to-end differentiable solution
• Relies on keypoint estimation to find the center points and regress all other object properties(such as size)
• As a result, the model is simpler, faster and more accurate than bounding-box based detectors

■Object as Points

■Training

■Keypoint Heatmap

■Local Offset

■Size Prediction

■Loss Function

■Network Architecture

• Authors experiment with different backbone architectures, obtaining different results:

Results without test augmentation(N.A.), flip testing(F), and multi-scale augmentation(MS). HW: Intel Core i7-8086k CPU, Titan Xp GPU

• The backbone that produces best speed/accuracy tradeoff is DLA-34 (modified by authors)

■Network Architecture

■Inference

■Results

■ダウンロード

Centernet.pdf