Even if the state-of-the-art SiamRPN tracker achieves very competitve results, as shown in [Li et al., 2018], this alternative architecture achieves near state-of-the-art performance on common benchmark datasets like OBT-2015 using no correlation map.
Link to the full project report: alexandre.carlier.page/projects/object_tracking
Success and Precision plots of SiamBroadcastRPN on OTB-2015
The SiamBroadcastRPN architecture. A MaxPool layer reduces the exemplar features to a vector, which can be concatenated to the search features after a Broadcast operation.
Similarly to SSD, bounding boxes are regressed from different feature maps.
To start training the network, run the command python3 train.py --config-file PATH/TO/CONFIG_FILE.
Configuration files are written in YAML (Yet Another Markup Language), which override the default configuration in configs/defaults.py.
To start benchmarking the model, run the command python3 benchmark.py --checkpoint PATH/TO/CHECKPOINT --config-file PATH/TO/CONFIG/FILE.