Pseudo-labelled boxes inversed in Soft Teacher

[eugene123tw]

Hi,

I've been doing some research in SSL, and found soft teacher implemented in mmdetection/3.x.

That being said, I have a question regarding the method get_pseudo_instances in Soft Teacher.

for data_samples, results in zip(batch_data_samples, results_list):
    data_samples.gt_instances = results

batch_data_samples = filter_gt_instances(batch_data_samples, score_thr=self.semi_train_cfg.pseudo_label_initial_score_thr)

reg_uncs_list = self.compute_uncertainty_with_aug(x, batch_data_samples)

for data_samples, reg_uncs in zip(batch_data_samples, reg_uncs_list):
    data_samples.gt_instances['reg_uncs'] = reg_uncs
    data_samples.gt_instances.bboxes = bbox_project(
        data_samples.gt_instances.bboxes,
        torch.from_numpy(data_samples.homography_matrix).inverse().to(self.data_preprocessor.device), 
        data_samples.ori_shape)

Why is it necessary to invert the pseudo-labelled boxes with bbox_project? Is it because the unsup student needs to apply its box projection?