Skip to content

Latest commit

 

History

History

NanoTrack

README.md

NanoTrack

【2023.02.25】Update NanoTrackV3

NanoTrackV1 Architecture

network

  • NanoTrack is a lightweight and high speed tracking network which mainly referring to SiamBAN and LightTrack. It is suitable for deployment on embedded or mobile devices. In fact, V1 and V2 can run at > 200FPS on Apple M1 CPU.
Trackers Backbone Size(*.onnx) Head Size (*.onnx) FLOPs Parameters
NanoTrackV1 752K 384K 75.6M 287.9K
NanoTrackV2 1.0M 712K 84.6M 334.1K
NanoTrackV3 1.4M 1.1M 115.6M 541.4K
  • Experiments show that NanoTrack has good performance on tracking datasets.
Trackers Backbone Model Size(*.pth) VOT2018 EAO VOT2019 EAO GOT-10k-Val AO GOT-10k-Val SR DTB70 Success DTB70 Precision
NanoTrackV1 MobileNetV3 2.4MB 0.311 0.247 0.604 0.724 0.532 0.727
NanoTrackV2 MobileNetV3 2.0MB 0.352 0.270 0.680 0.817 0.584 0.753
NanoTrackV3 MobileNetV3 3.4MB 0.449 0.296 0.719 0.848 0.628 0.815
CVPR2021 LightTrack MobileNetV3 7.7MB 0.418 0.328 0.75 0.877 0.591 0.766
WACV2022 SiamTPN ShuffleNetV2 62.2MB 0.191 0.209 0.728 0.865 0.572 0.728
ICRA2021 SiamAPN AlexNet 118.7MB 0.248 0.235 0.622 0.708 0.585 0.786
IROS2021 SiamAPN++ AlexNet 187MB 0.268 0.234 0.635 0.73 0.594 0.791

  • For NanoTrackV1, we provide Android demo and MacOS demo based on ncnn inference framework.

  • We also provide PyTorch code. It is friendly for training with much lower GPU memory cost than other models. NanoTrackV1 only uses GOT-10k dataset to train, which only takes two hours on RTX3090.

PyTorch

  • Build
python  setup.py build_ext --inplace
  • Prepare data
1. cd xxx/xxx/NanoTrack 

2. mkdir data
Download GOT-10k https://pan.baidu.com/s/10crE2uKR182fA93XRB3jyg password: 5ebm
How to crop GOT-10k https://pan.baidu.com/s/1yzXnRUqjreSeojEQOmluQg  password: un47
Put your training data into data directory 

3. mkdir datasets
Download VOT2018 https://pan.baidu.com/s/1MOWZ5lcxfF0wsgSuj5g4Yw password: e5eh  
Put your testing data into datasets directory
  • Select NanoTrackV1
file: ./nanotrack/models/head/__init__.py
    from nanotrack.models.head.ban_v1 import UPChannelBAN, DepthwiseBAN

file: ./models/config/configv1.yaml
    WINDOW_INFLUENCE: 0.462
    PENALTY_K: 0.148 
    LR: 0.390
  • Select NanoTrackV2
file: ./nanotrack/models/head/__init__.py
    from nanotrack.models.head.ban_v2 import UPChannelBAN, DepthwiseBAN

file: ./models/config/configv2.yaml
    WINDOW_INFLUENCE: 0.490
    PENALTY_K: 0.150
    LR: 0.385
  • Select NanoTrackV3
file: ./nanotrack/models/head/__init__.py
    from nanotrack.models.head.ban_v3 import UPChannelBAN, DepthwiseBAN

file: ./models/config/configv3.yaml
    WINDOW_INFLUENCE: 0.455
    PENALTY_K: 0.138
    LR: 0.348
  • Train
python ./bin/train.py
  • Test
python ./bin/test.py
  • Eval
python ./bin/eval.py
  • Search params
python ./bin/hp_search.py 

python ./bin/eval.py \
--tracker_path ./hp_search_result \
--dataset VOT2018 \
--num 4 \
--tracker_name  'checkpoint*'
  • Calculate flops
python cal_macs_params.py
  • Calculate speed
python cal_speed.py
  • PyTorch to ONNX
python ./pytorch2onnx.py
  • ONNX to NCNN
https://convertmodel.com/

MacOS

PC demo

  • Modify CMakeList.txt

  • Build (Apple M1 CPU)

    $ sh make_macos_arm64.sh

Android

Android demo

  • Modify CMakeList.txt

  • Download(password: 6cdd) OpenCV and NCNN libraries for Android ◊

Reference