LightTrack: Finding Lightweight Neural Networks for Object Tracking via One-Shot Architecture Search

The official implementation of the paper

LightTrack: Finding Lightweight Neural Networks for Object Tracking via One-Shot Architecture Search

Hiring research interns for visual transformer projects: houwen.peng@microsoft.com

News

We have uploaded the pre-trained weights of the SuperNets(for both ImageNet classification and object tracking) to Google Drive. Users can use them as initialization for future research on efficient object tracking.

Abstract

We present LightTrack, which uses neural architecture search (NAS) to design more lightweight and efficient object trackers. Comprehensive experiments show that our LightTrack is effective. It can find trackers that achieve superior performance compared to handcrafted SOTA trackers, such as SiamRPN++ and Ocean, while using much fewer model Flops and parameters. Moreover, when deployed on resource-constrained mobile chipsets, the discovered trackers run much faster. For example, on Snapdragon 845 Adreno GPU, LightTrack runs 12× faster than Ocean, while using 13× fewer parameters and 38× fewer Flops. Such improvements might narrow the gap between academic models and industrial deployments in object tracking task.

Environment Installation

cd lighttrack
conda create -n lighttrack python=3.6
conda activate lighttrack
bash install.sh

Data Preparation

Tracking Benchmarks

Please put VOT2019 dataset under $LightTrack/dataset. The prepared data should look like:

$LighTrack/dataset/VOT2019.json
$LighTrack/dataset/VOT2019/agility
$LighTrack/dataset/VOT2019/ants1
...
$LighTrack/dataset/VOT2019/list.txt

Test and evaluation

Test LightTrack-Mobile on VOT2019

bash tracking/reproduce_vot2019.sh

Flops, Params, and Speed

Compute the flops and params of our LightTrack-Mobile. The flops counter we use is pytorch-OpCounter

python tracking/FLOPs_Params.py

Test the running speed of our LightTrack-Mobile

python tracking/Speed.py