Learning Object-aware Anchor-free Networks for Real-time Object Tracking

Code and models for double blind review in ECCV2020.

Introduction

Anchor-based Siamese trackers have witnessed tremendous success in visual tracking, yet their robustness still leaves a gap with recent state-of-the-art approaches. We find the underlying reason is that the regression network in anchor-based methods is only trained on the positive anchor boxes. This mechanism makes it difficult to refine the anchors whose overlap with the target objects are small.

Our proposals mitigate the problem by,

We propose a novel anchor-free algorithm for realtime visual tracking. It is capable of rectifying the imprecise bounding-box predictions whose overlap with the target objects are small.
We introduce an adaptive feature alignment mechanism to set up a correspondence between the objectaware features and the predicted bounding boxes. This leads to a more robust classification of foreground objects and background clutters.

Quick Start

Installation

bash install.sh ocean

Preparation

The test datasets should be arranged in dataset directory, and the pre-trained models should be arranged in snapshot directory.

${Tracking_ROOT}
|—— experimnets
|—— lib
|—— snapshot
  |—— xxx.model
|—— dataset
  |—— VOT2019.json 
  |—— VOT2019
     |—— videos...
|—— ...

Download pre-trained models and json files here.

Test

python ocean_tracking/test_ocean.py --arch Ocean --dataset VOT2018

Evaluation

python lib/eval_toolkit/bin/eval.py --dataset_dir dataset --dataset VOT2018 --tracker_result_dir result/VOT2018 --trackers Ocean

Train

Please follow the tutorial here.

wangaixue/ocean