siamrpn.pytorch

This repository is clean re-implement of siamrpn using PyTorch that is more readable for newers. 这个repo就看看就行，主要是为了简便看懂siamrpn设计的，有些人反馈lmbd读取有问题不能训练，之后有同样问题的就不用反馈了

Environment

• Ubuntu 18.04
• PyTorch1.2
• CUDA10.0 + CUDNN7.4.2 (not so strict)
• TO install needed packages: pip install -r requirements.txt

Prepare dataset

There are two ways to prepare datasets:

1. This is the easier one. Just download my prepared three files(two .mdb and one .json) at link1, link2 and link3, and put them all in dataset directory
2. This is the another one: refer to pysot dataset part and prepare the YOUTUBEBB and VID. Then, you need modify some path items in config.config.py like me:

__C.DATASET.VID.ROOT = '/home/lz/Videos/VID/crop511'
__C.DATASET.VID.ANNO = '/home/lz/PycharmProjects/pysot-master/training_dataset/vid/train.json'
__C.DATASET.YOUTUBEBB.ROOT = '/home/lz/Videos/yt_bb/crop511'
__C.DATASET.YOUTUBEBB.ANNO = '/home/lz/PycharmProjects/pysot-master/training_dataset/yt_bb/train.json'

and cd dataset run python prepare_dataset.py, after 1.5 hours you will get ready like first way.

Train

1. cd project root directory
2. you can modify some items, such as BATCH_SIZE,NUM_WORKERS in config/config.yaml
3. then run

python train.py

Test

This repo use got-10k toolkit to evaluate performance in OTB benchmark, so you need to do follow things:

1. Download the raw OTB dataset and unzip all videos(otherwise will download and unzip automatically)
2. run

python test_OTB.py --root_dir='your_OTB_dir_in_step1'

Result

I have just make it work and need more works to train well, so the result will not provide including pretrained model and OTB100 performance now. See TODO

TODO

• train and test phase
• hyper-parameters search
• multi-GPU training
• higher performance