how to use PURA

[saymamba]

Congratulations! Excellent work!
This paper focuses on the inference stage. Could you please explain how it's used specifically? Do we need to train PURA during training? Has this part of the code been released yet?

[MelanTech]

Congratulations! Excellent work! This paper focuses on the inference stage. Could you please explain how it's used specifically? Do we need to train PURA during training? Has this part of the code been released yet?

PURA is an unsupervised test-time adaptation method that only takes effect during the inference phase and does not require additional training. For trackers developed based on the pytracing framework, you can enable PURA by following these steps. Firstly, download the pura.py file from this code repository and copy it to the lib\test\tracker directory. Secondly, add two lines of code to the lib\test\tracker\xxx_track.py to complete the integration.

from lib.test.tracker import pura  # import PURA

...

class XXXTrack(BaseTracker):
    def __init__(self, params, dataset_name):
        super(XXXTrack, self).__init__(params)
        network = build_xxx_track(params.cfg, training=False)
        network.load_state_dict(torch.load(self.params.checkpoint, map_location='cpu')['net'], strict=True)
        self.cfg = params.cfg
        self.network = network.cuda()
        self.network.eval()

        # Enable PURA
        pura.replace_batchnorm(self.network.box_head)  # replace batchnorm in box_head with PURA
        self.network = pura.configure_model(self.network)  # configure model

        self.preprocessor = Preprocessor()
        self.state = None
        ...

Currently, our baseline code is under review. Premature release of code may have a negative impact on the review process. Therefore, at this stage, we will only release the core code of PURA. PURA is compatible with all trackers built on the pytracking framework. After the review process is completed, we will immediately release the complete code to facilitate your experimental replication.