Convolutional Neural Networks for Visual Tracking 2015 02

Convolutional Neural Networks for Visual Tracking 2015. 02. 17 Computer Vision Lab. 남현섭

Contents • H. Li, Y. Li, and F. Robust Online Visual Tracking with an Single Convolutional Neural Network, ACCV 2014 • N. Wang, S. Li, A. Gupta and D. Y. Yeung, Transferring Rich Feature Hierarchies for Robust Visual Tracking, ar. Xiv 2015

H. Li, Y. Li, and F. Robust Online Visual Tracking with an Single Convolutional Neural Network ACCV 2014

Contributions • Online learning of a single CNN • Robust, structural, truncated loss function • Iterative SGD method with a temporal sampling

CNN Architecture

Structural Loss Function • Increase the number of available training samples by introducing the structural importance • Traditional loss function • Structural loss function Structural importance CNN loss overlapping ratio

Structural Loss With a Robust Term • Regularize the ordinary structural loss using a set of positive instances rather than only one positive instance.

Structural Loss With a Robust Term • Predicted state • Positive instance set • Robust term • Structural loss function with a robust term

Structural Loss With a Robust Term and the Truncated Norm •

Online Learning of CNN • Temporal Sampling – Learn a discriminative model on a long-term positive set and a shortterm negative set. • Iterative Stochastic Gradient Descent – Iteratively update the convolutional layers, and jointly update the fully connected layers.

Online Learning of CNN

Online Learning of CNN

Results

Results (CNN tracker variations)

N. Wang, S. Li, A. Gupta and D. Y. Yeung Transferring Rich Feature Hierarchies for Robust Visual Tracking ar. Xiv 2015

Contributions • Transferring a pretrained CNN into online tracking • Structured output corresponding to a probability map

Objectness Pretraining • Pretrain a CNN to learn generic features for distinguishing objects from non-objects

Online Tracking – State Determination – Sample Collection – Differentially-paced Fine-tuning

Online Tracking • State Determination average

Online Tracking •

Results

Results (non-rigid object)

Discussion • Transferring a pretrained CNN • Employing structural information • Modifying loss function • Efficient online learning • CNN update policy

On-going research Finetuned fully connected layers Pretrained convolutional layers conv+pool 3 96 conv+pool 256 conv+pool conv 384 256 64 2 Probability map 256 64 2