Skip to content

Latest commit

 

History

History
80 lines (59 loc) · 3.79 KB

ocean.md

File metadata and controls

80 lines (59 loc) · 3.79 KB

Ocean tutorial

Testing

We assume the root path is $SOTS, e.g. /home/zpzhang/SOTS

Set up environment

cd $SOTS/lib/tutorial
bash install.sh $conda_path SOTS
cd $SOTS
conda activate SOTS
python setup.py develop

$conda_path denotes your anaconda path, e.g. /home/zpzhang/anaconda3

  • [Optional] Install TensorRT according to the tutorial.

Note: we perform TensorRT evaluation on RTX2080 Ti and CUDA10.0. If you fail to install it, please use pytorch version.

Prepare data and models

  1. Download the pretrained PyTorch model and TensorRT model to $SOTS/snapshot.
  2. Download json files of testing data and put them in $SOTS/dataset.
  3. Download testing data e.g. VOT2019 and put them in $SOTS/dataset. Please download each data from their official websites, and the directories should be named like VOT2019, OTB2015, GOT10K, LASOT.

Testing

In root path $SOTS,

python tracking/test_ocean.py --arch Ocean --resume snapshot/OceanV.pth --dataset VOT2019

Evaluation

python lib/eval_toolkit/bin/eval.py --dataset_dir dataset --dataset VOT2019 --tracker_result_dir result/VOT2019 --trackers Ocean

You may test other datasets with our code. Please corresponds the provided pre-trained model --resume and dataset --dataset. See ocean_model.txt for their correspondences.

TensorRT toy

Testing video: twinnings in OTB2015 (472 frames) Testing GPU: RTX2080Ti

  • TensorRT (149fps)
python tracking/test_ocean.py --arch OceanTRT --resume snapshot/OceanV.pth --dataset OTB2015 --video twinnings
  • Pytorch (68fps)
python tracking/test_ocean.py --arch Ocean --resume snapshot/OceanV.pth --dataset OTB2015 --video twinnings

Note:

  • TensorRT version of Ocean only supports 255 input.
  • Current TensorRT does not well support some operations. We would continuously renew it following official TensorRT updating. If you want to test on the benchmark, please us the Pytorch version.
  • If you want to use our code in a realistic product, our TensorRT code may help you.

☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️☁️

Training

prepare data

  • Please download training data from GoogleDrive or BaiduDrive(urxq), and then put them in $SOTS/data
  • You could also refer to scripts in $SOTS/lib/dataset/crop to process your custom data.
  • For splited files in BaiduDrive, please use cat got10k.tar.* | tar -zxv to merge and unzip.

prepare pretrained model

Please download the pretrained model on ImageNet here, and then put it in $SOTS/pretrain.

modify settings

Please modify the training settings in $SOTS/experiments/train/Ocean.yaml. The default number of GPU and batch size in paper are 8 and 32 respectively.

run

In root path $SOTS,

python tracking/onekey.py

This script integrates train, epoch test and tune. It is suggested to run them one by one when you are not familiar with our whole framework (modify the key ISTRUE in $SOTS/experiments/train/Ocean.yaml). When you know this framework well, simply run this one-key script.