The official implementation of LightFC
Option1: Use the Anaconda
conda create -n lightfc python=3.9
conda activate lightfc
bash install.sh
Follow stark and ostrack frameworks to set your datasets
Project file directory should be like
${YOUR_PROJECT_ROOT}
-- experiments
|-- lightfc
-- external
|-- vot20st
-- lib
|--models
...
-- outputs (download and unzip the output.zip to obtain our checkpoints and row results)
|--checkpoints
|--...
|--test
|--...
-- pretrained_models (if you want to train lightfc, put pretrained model here)
|--mobilenetv2.pth (from torchvision model)
...
-- tracking
...
Download lightfc checkpoint and raw results at Google Drive
Download lightfc-vit checkpoint and raw results at Google Drive
Then go to these two files, and modify the paths
lib/train/admin/local.py # paths about training
lib/test/evaluation/local.py # paths about testing
Training with multiple GPUs using DDP
python tracking/train.py --script LightFC --config mobilnetv2_p_pwcorr_se_scf_sc_iab_sc_adj_concat_repn33_se_conv33_center_wiou --save_dir . --mode multiple --nproc_per_node 2
If you want to train lightfc, please download https://download.pytorch.org/models/mobilenet_v2-b0353104.pth rather than https://download.pytorch.org/models/mobilenet_v2-7ebf99e0.pth
if you want to train lightfc-vit, please download https://github.com/wkcn/TinyViT-model-zoo/releases/download/checkpoints/tiny_vit_5m_22k_distill.pth
Go to tracking/test.py and modify the parameters
python tracking/test.py
Then go to tracking/analysis_results.py and modify the parameters
python tracking/analysis_results.py
# Params and FLOPs
python tracking/profile_model.py
# Speed
python tracking/speed.py