Implentations of some deep networks on GOT-10K.
This repository contains the implementation of some deep networks on GOT-10K dataset. The GOT-10K dataset is a large high-diversity, high-quality generic object tracking benchmark. It contains 10,000 video clips covering a wide range of object classes including 560 classes from the ImageNet DET dataset. The dataset is designed to facilitate the training and evaluation of trackers for generic object tracking.
The implemented models are:
- SiamFC: Fully-Convolutional Siamese Networks for Object Tracking.
- SiamRPN: High Performance Visual Tracking with Siamese Region Proposal Network.
- SiamDW: Deeper and Wider Siamese Networks for Real-Time Visual Tracking.
- ARTrack: Adaptive and Robust Tracking with Siamese Networks.
- MixFormer: End-to-End Tracking with Iterative Mixed Attention
- SiamFamilies: A collection of SiamFC, SiamRPN, SiamDW, etc.
- Python 3.6
- PyTorch 1.0.0
- torchvision 0.2.1
- Download the GOT-10K dataset from here.
- Extract the dataset to the
data
folder. - Run the
train.py
script to train the model. - Run the
test.py
script to test the model.
The structure of the project is as follows:
.
├── data
│ ├── GOT-10k
│ │ ├── train
│ │ │ ├── GOT-10k_Train_000001
│ │ │ │ ├── 000000.jpg
│ │ │ │ ├── 000001.jpg
│ │ │ │ ├── ...
│ │ │ ├── ...
│ │ ├── val
│ │ │ ├── GOT-10k_Val_000001
│ │ │ │ ├── 000000.jpg
│ │ │ │ ├── 000001.jpg
│ │ │ │ ├── ...
│ │ │ ├── ...
├── models
│ ├── siamfc.py
│ ├── siamrpn.py
│ ├── siamdw.py
│ ├── artrack.py
├── train.py
├── test.py
The following table shows the results of the implemented models on the GOT-10K dataset.
Model | AO | SR_0.50 | SR_0.75 | Hz |
---|---|---|---|---|
SiamFC | 0.348 | 0.353 | 0.098 | 44.15 fps |
SiamRPN | 0.378 | 0.384 | 0.116 | 35.71 fps |
SiamDW | 0.394 | 0.400 | 0.124 | 30.30 fps |
ARTrack | 0.412 | 0.418 | 0.134 | 27.03 fps |
MixFormer | 0.426 | 0.432 | 0.142 | 24.39 fps |