--MDNet: Real-Time Multi-Domain Convolutional Neural Network Tracker Created by Ilchae Jung, Jeany Son, Mooyeol Baek, and Bohyung Han
- python>=3
- pytorch>=1.0
- some others library functions
For more detailed packages, refer to [MDNet](https://github.com/hyeonseobnam/py-MDNet) .
In our tracker, we use MDNet as our backbone and extend to multi-modal tracker.We use imagenet-vid.pth as our pretrain model.Then we use this with the training model in GTOT and RGBT234 models to pre-train our dual-stream MDNet_RGBT backbone network.And thus we get the GTOT.pth and RGBT234.pth.And Then We load the basic model to
train Our network and get the final model.Our model and the pretrain model is available at pth model.The extract code is mmic.After downloading the model, you should put it in AAAICode/models/
In the tracking/Run.py file, you need to change dataset path, model_path and result_path In the tracking/Run.py file. You can load the model GTOT_ALL_Transformer for testing RGBT234 and LasHeR. And use the RGBT234_ALL_Transformer for testing the GTOT.
python ./tracking/Run.py
There Stage train:
At First you should use the GTOT and RGBT234 datasets with challenge tags,you can find the datasets in (https://github.com/mmic-lcl/Datasets-and-benchmark-code) run prepro_data.py generate a xxx.pkl file to store the data path.Please note that you should adjust the phased training parameters in the pretrain_option.py when training each stage.
- In the first stage,you should run the train_stage1.py 5 times because we have five attribute branches. Each time we train the network with specific label data, we load the pre-trained backbone network model parameters and then add the specific branch one by one for training. Note that at this stage we only save each branch model parameters.
- You have spawned 5 corresponding challenge branches in one phase.In the second stage,you can load the backbone and all branches parameters for training the Attribute-Based Aggregation Fusion model. you should run the train_stage2.py
- On the basis of the two-stage, only the model parameters generated by the second stage need to be loaded, and the final model can be generated in the three-stage training. you should run the train_stage3.py