MasterBin-IIAU
commented
2 years ago @orilifs Hi, thanks for your appreciation of our work.
Answer to Q1.a: These two models are part of our improved version of STARK (we call it STARK-Lightning). The first one (backbone_bottleneck_pe.py) is the template branch, which only need to run one-time (on the first frame). The second one (complete.py) is the search region branch, which takes the search image, the extracted template feature as the input, and need to run on each frame.
Answer to Q1.b: The details about STARK-Lightning have not been covered in the original paper of STARK. We may consider adding these parts in later version. For now you can refer to the PPT of STARK. (https://github.com/researchmm/Stark/blob/main/external/STARK-ExtreMart.pdf)
Answer to Q2: The config file "baseline_rephead_4_lite_search5.yaml" provides details about STARK-Lightning (training settings, model architectures, etc). You can find that we build the STARK-Lightning model based on this config. (https://github.com/researchmm/Stark/blob/5daa6f0e939727c68ae13f273cf1ef7511f4ec9a/tracking/ORT_lightning_X_trt_complete.py#L88). Besides, the previously mentioned two models are only used when you want to run STARK-Lightning model with the ONNXRuntime.
Answer to Q2.b: For now there is only one config file about STARK-Lightning. We may consider adding more config files for supporting more backbones for example.
orilifs
Hello, thank you for your great work! Amazing results. @MasterBin-IIAU @penghouwen @jb2020-super
1.a. What is the difference between these two models? python tracking/ORT_lightning_X_trt_backbone_bottleneck_pe.py # for the template branch python tracking/ORT_lightning_X_trt_complete.py # for the search region branch
1.b. Where can I read about these two models? Is it in the paper?
2.b. Are there other config files for the lightning_Xtrt**** models please?