where the GEM module?

[caiyujue]

I find this code:"dist_map = self.create_dist(init_patch_crop.shape[0], init_patch_crop.shape[1])",but is that the "The target localisation channel (L)" mentioned in the paper? In segm_net.py the code : " segm_layers = torch.cat((torch.unsqueeze(pred_sm[:, :, :, 0], dim=1), torch.unsqueeze(pred_pos, dim=1), dist), dim=1)", It looks like this "The target localisation channel (L)" is fixed as the code "self.dist_map = dist_map",has the "self.dist_map" changed?and where ? thanks~

[caiyujue]

maybe I should change the " params.segm_dist_map_type = 'center' " to params.segm_dist_map_type = 'bbox'?

[caiyujue]

I changed the " params.segm_dist_map_type = 'center' ,but the performance has deteriorated in VOT2018,when it's center mode,lost_num=47,while bbox mode ,the lost_num = 132,and I test the tracker with the pre-trained model also encountered the same problem.

[alanlukezic]

The segm_dist_map_type parameter should be set to 'center' since the network was trained this way. During training the L channel is simulated from the ground-truth target position, while the inference is implemented so that the L channel remains constant since target is cropped on the position estimated with GEM.

[caiyujue]

thank you for your reply

[laisimiao]

@caiyujue Duplicate of #21