ICCV 2015 Tutorial on Tools for Efficient Object Detection

December 11, 2015
ICCV 2015, Santiago, Chile


Rogerio Feris
IBM Research
Piotr Dollar
Facebook AI Research
Xiaoyu Wang
Snapchat Research
Kaiming He
Microsoft Research
Ross Girshick
Facebook AI Research
Rodrigo Benenson
Max Planck Institute
Jan Hosang
Max Planck Institute

Tutorial description

In many real-world applications, running a fast object detector is as critical as running an accurate object detector. Significant progress has been made in the past few years to boost th e accuracy levels of both image classification and object detection, but existing solutions often rely on computationally expensive models, which are prohibitively slow for numerous applica tions. Building a fast object detector is also challenging due to the large number of image sub-windows that usually need to be scanned in order to localize objects in the image. The goal o f this tutorial is to present a set of modern tools for efficient and accurate object detection. We will cover emerging topics such as region proposals, fast feature pyramids, and state-of- the-art detectors based on fast extraction of Convolutional Neural Network (CNN) features. The organizers will also share their experience in building real-time systems through hands-on ses sions based on publicly available source code.


Time Lecture Topics


(Rogerio Feris, pdf slides)
  • Motivation, applications, and course overview

Detecting Objects at 100 Hz with Rigid Templates

(Rodrigo Benenson, pdf slides )
  • Feature channels
  • Approximate feature pyramids
  • Hard, soft, and cross-talk cascades
  • ICF Detector and extensions
  • Case studies: real-time pedestrian and face detection
15:00 Coffee Break

Region Proposals

(Jan Hosang, pdf slides )
  • Grouping proposal methods
  • Window scoring proposal methods
  • Metrics and in-depth analysis

Regionlet Object Detector with Hand-crafted and CNN Features

(Xiaoyu Wang, pdf slides)
  • Fast Object Detection with Regionlets
  • Regionlets and Deep Neural Networks Integration

Convolutional Feature Maps: Elements of Efficient (and Accurate) CNN-based Object Detection

(Kaiming He, pdf slides)
  • From HOG feature maps to conv feature maps
  • From image regions to feature map regions
  • Adaptive pooling (SPP/RoI) on feature maps
  • For detection (SPPnet, Fast R-CNN) and region proposals (RPN)

Training R-CNNs of Various Velocities: Slow, Fast, and Faster

(Ross Girshick, pdf slides , pptx slides)
  • Motivation for training very deep models end-to-end
  • Training with SPP/RoI pooling layers (bprop, image-centric sampling)
  • Fast R-CNN
  • Faster R-CNN
    • Alternating optimization
    • Approximate joint training with RPN
  • Code pointers
18:00 Concluding Remarks