Visual Tracking by Means of Deep Reinforcement Learning and an Expert Demonstrator

Abstract

In the last decade many different algorithms have been proposed to track a generic object in videos. Their execution on recent large-scale video datasets can produce a great amount of various tracking behaviours. New trends in Reinforcement Learning showed that demonstrations of an expert agent can be efficiently used to speed-up the process of policy learning. Taking inspiration from such works and from the recent applications of Reinforcement Learning to visual tracking, we propose two novel trackers, A3CT, which exploits demonstrations of a state-of-the-art tracker to learn an effective tracking policy, and A3CTD, that takes advantage of the same expert tracker to correct its behaviour during tracking. Through an extensive experimental validation on the GOT-10k, OTB-100, LaSOT, UAV123 and VOT benchmarks, we show that the proposed trackers achieve state-of-the-art performance while running in real-time. arXiv:1909.08487v1 [cs.CV] 18 Sep 2019 =

Weakly-Supervised Domain Adaptation of Deep Regression Trackers via Reinforced Knowledge Distillation

“…Tracker. We follow the most recent advancements in deep regression tracking [7], [34], [9] to implement our tracker s(•|θ) as a deep neural network with weights θ. The network gets as input s t as two image patches which pass through two ResNet-18 [35] CNN branches with shared weights.…”

Deep Reinforcement Learning in Computer Vision: A Comprehensive Survey

“…An improvement to [45] for SOT was proposed by [84], where a visual tracker was formulated using DRL and an expert demonstrator. The authors treated the problem as an MDP, where the state consists of two consecutive frames that have been cropped using the bounding box corresponding to the former frame and used a scaling factor to control the offset while cropping.…”

“…The figure illustrates a general implementation of object tracking in videos using DRL, where the state consists of two consecutive frames (F t , F t+1 ) with a bounding box for the first frame produced by another algorithm for the first iteration or by the previous iterations of DRL agent. The actions corresponds to the moving the bounding on the image to fit the object in frame F t+1 , hence forming a new state with frame F t+1 and frame F t+2 along with the bounding box for frame F t+1 predicted by previous iteration and reward corresponds to whether IOU is greater then a given threshold as used by [118], [308], [45], [84], [307], [168], [169].…”

Video-Based Reconstruction of the Trajectories Performed by Skiers

“…< l a t e x i t s h a 1 _ b a s e 6 4 = " 0 q a i s a C W N S A d B x M e B n 5 Athlete Tracking. The first step of the pipeline is to ploit a visual object tracker [9,10] to track the motion of the athlete across all the frames up to F t . We used a tracker outputting a bounding-box b t = (x The number in the top-left corner of each image reports the frame index t in the video.…”