Number of unique track ids in dataset zoo

[youonlytrackonce]

Hello,

As the title states, I have been investigating the numbers of unique track ids in data_all dataset. Because there are more than one dataset and each may have more than one sequence, how can the network learn reid embeddings for each unique ids? There can be colliding track ids in the different datasets.

[janthmueller]

Regarding the subsets of a dataset, the track ids already exist in this format, or are created using the gen_label modules. Regarding multiple datasets, for example, within the JointDataset class these ids are processed one more time so that the maximum id of the preceding datasets is added to all ids of the following dataset.

[sompt22]

Regarding the subsets of a dataset, the track ids already exist in this format, or are created using the gen_label modules. Regarding multiple datasets, for example, within the JointDataset class these ids are processed one more time so that the maximum id of the preceding datasets is added to all ids of the following dataset.

Could you please point the file and line of the code where maximum id of the preceding datasets is added to all ids of the following dataset? Is it on the fly process or applied on the disk?

[janthmueller]

Regarding the subsets of a dataset, the track ids already exist in this format, or are created using the gen_label modules. Regarding multiple datasets, for example, within the JointDataset class these ids are processed one more time so that the maximum id of the preceding datasets is added to all ids of the following dataset.

Could you please point the file and line of the code where maximum id of the preceding datasets is added to all ids of the following dataset? Is it on the fly process or applied on the disk?

It is applied on the fly. See the JointDataset class (in .\src\lib\datasets\dataset\jde.py) at the beginning of the __getitem__ method (line 426-428):

        for i, _ in enumerate(labels):
            if labels[i, 1] > -1:
                labels[i, 1] += self.tid_start_index[ds]

The self.tid_start_index attribute gets computed in the __init__ (line 524-527 with tid_num computed just before that).